Artificial Intelligence (AI)
is a commonly employed appellation to refer to the field of science aimed at providing machines with the capacity of performing functions such as logic, reasoning, planning, learning, and perception. Despite the reference to “machines” in this definition, the latter could be applied to “any type of living intelligence”. Likewise, the meaning of intelligence, as it is found in primates and other exceptional animals for example, it can be extended to include an interleaved set of capacities, including creativity, emotional knowledge, and self-awareness.
The term AI was closely associated with the field of “symbolic AI”, which was popular until the end of the 1980s. In order to overcome some of the limitations of symbolic AI, subsymbolic methodologies such as neural networks, fuzzy systems, evolutionary computation and other computational models started gaining popularity, leading to the term “computational intelligence” emerging as a subfield of AI.
Nowadays, the term AI encompasses the whole conceptualisation of a machine that is intelligent in terms of both operational and social consequences. A practical definition used is one proposed by Russell and Norvig: “Artificial Intelligence is the study of human intelligence and actions replicated artificially, such that the resultant bears to its design a reasonable level of rationality” [1]. This definition can be further refined by stipulating that the level of rationality may even supersede humans, for specific and well-defined tasks.
Current AI technologies are used in online advertising, driving, aviation, medicine and personal assistance image recognition. The recent success of AI has captured the imagination of both the scientific community and the public. An example of this is vehicles equipped with an automatic steering system, also known as autonomous cars. Each vehicle is equipped with a series of lidar sensors and cameras which enable recognition of its three-dimensional environment and provides the ability to make intelligent decisions on maneuvers in variable, real-traffic road conditions. Another example is the Alpha-Go, developed by Google Deepmind, to play the board game Go. Last year, Alpha-Go defeated the Korean grandmaster Lee Sedol, becoming the first machine to beat a professional player and recently it went on to win against the current world number one, Ke Jie, in China. The number of possible games in Go is estimated to be 10761 and given the extreme complexity of the game, most AI researchers believed it would be years before this could happen. This has led to both the excitement and fear in many that AI will surpass humans in all the fields it marches into.
However, current AI technologies are limited to very specific applications. One limitation of AI, for example, is the lack of “common sense”; the ability to judge information beyond its acquired knowledge. A recent example is that of the AI robot Tay developed by Microsoft and designed for making conversations on social networks. It had to be disconnected shortly after its launch because it was not able to distinguish between positive and negative human interaction. AI is also limited in terms of emotional intelligence. AI can only detect basic human emotional states such as anger, joy, sadness, fear, pain, stress and neutrality. Emotional intelligence is one of the next frontiers of higher levels of personalisation.
True and complete AI does not yet exist. At this level, AI will mimic human cognition to a point that it will enable the ability to dream, think, feel emotions and have own goals. Although there is no evidence yet this kind of true AI could exist before 2050, nevertheless the computer science principles driving AI forward, are rapidly advancing and it is important to assess its impact, not only from a technological standpoint, but also from a social, ethical and legal perspective.
THE BIRTH AND BOOM OF AI
The birth of the computer took place when the first calculator machines were developed, from the mechanical calculator of Babbage, to the electromechanical calculator of Torres-Quevedo. The dawn of automata theory can be traced back to World War II with what was known as the “codebreakers”. The amount of operations required to decode the German trigrams of the Enigma machine, without knowing the rotor’s position, proved to be too challenging to be solved manually. The inclusion of automata theory in computing conceived the first logical machines to account for operations such as generating, codifying, storing and using information. Indeed, these four tasks are the basic operations of information processing performed by humans. The pioneering work by Ramón y Cajal marked the birth of neuroscience, although many neurological structures and stimulus responses were already known and studied before him. For the first time in history the concept of “neuron” was proposed. McClulloch and Pitts further developed a connection between automata theory and neuroscience, proposing the first artificial neuron which, years later, gave rise to the first computational intelligence algorithm, namely “the perceptron”. This idea generated great interest among prominent scientists of the time, such as Von Neumann, who was the pioneer of modern computers and set the foundation for the connectionism movement.
In 1971, DARPA funded a consortium of leading laboratories in the field of speech recognition. The project had the ambitious goal of creating a fully functional speech recognition system with a large vocabulary. In the middle of the 1970s, the field of AI endured fierce criticism and budgetary restrictions, as AI research development did not match the overwhelming expectations of researchers.. When promised results did not materialize, investment in AI eroded. Following disappointing results, DARPA withdrew funding in speech recognition and this, coupled with other events such as the failure of machine translation, the abandonment of connectionism and the Lighthill report, marked the first winter of AI [2]. During this period, connectionism stagnated for the next 10 years following a devastating critique by Marvin Minksy on perceptrons [3].
From 1980 until 1987, AI programmes, called “expert systems”, were adopted by companies and knowledge acquisition become the central focus of AI research. At the same time, the Japanese government launched a massive funding program on AI, with its fifth-generation computers initiative. Connectionism was also revived by the work of John Hopfield [4] and David Rumelhart [5].
AI researchers who had experienced the first backlash in collapsed and billions of dollars were lost. The difficulty of updating and reprograming the expert systems, in addition to the high maintenance costs, led to the second AI winter. Investment in AI dropped and DARPA stopped its strategic computing initiative, claiming AI was no longer the “latest mode”. Japan also stopped funding its fifth-generation computer program as the proposed goals were not achieved.
In the 1990s, the new concept of “intelligent agent” emerged [6]. An agent is a system that perceives its environment
and undertakes actions that maximize its chances of being successful. The concept of agents conveys, for the first time, the idea of intelligent units working collaboratively with a common objective. This new paradigm was intended to mimic how humans work collectively in groups, organizations and/or societies. Intelligent agents proved to be a more polyvalent concept of intelligence. In the late 1990s, fields such as statistical learning from several perspectives including probabilistic, frequentist and possibilistic (fuzzy logic) approaches, were linked to AI to deal with the uncertainty of decisions. This brought a new wave of successful applications for AI, beyond what expert systems had achieved during the 1980s. These new ways of reasoning were more suited to cope with the uncertainty of intelligent agent states and perceptions and had its major impact in the field of control. During this time, high-speed trains controlled by fuzzy logic, were developed [7] as were many other industrial applications (e.g. factory valves, gas and petrol tanks surveillance, automatic gear transmission systems and reactor control in power plants) as well as household appliances with advanced levels of intelligence (e.g. air-conditioners, heating systems, cookers and vacuum- cleaners). These were different to the expert systems in 1980s; the modelling of the inference system for the task, achieved through learning, gave rise to the field of Machine Learning. Nevertheless, although machine reasoning exhibited good performance, there was still an engineering requirement to digest the input space into a new source, so that intelligence could reason more effectively. Since 2000, a third renaissance of the connectionism paradigm arrived with the dawn of Big Data, propelled by the rapid adoption of the Internet and mobile communication. Neural networks were once more considered, particularly in the role they played in enhancing perceptual intelligence and eliminating the necessity of feature engineering. Great advances were also made in computer vision, improving visual perception, increasing the capabilities of intelligent agents and robots in performing more complex tasks, combined with visual pattern recognition. All these paved the way to new AI challenges such as, speech recognition, natural language processing, and self-driving cars. A timeline of key highlights in the history of AI is shown in Figure 1.
1. QUESTIONING THE IMPACT OF AI
Given the exponential rise of interest in AI, experts have called for major studies on the impact of AI on our society, not only in technological but also in legal, ethical and socio- economic areas. This response also includes the speculation that autonomous super artificial intelligence may one day supersede the cognitive capabilities of humans. This future scenario is usually known in AI forums as the “AI singularity” [8]. This is commonly defined as the ability of machines to build better machines by themselves. This futuristic scenario has been questioned and is received with scepticism by many experts. Today’s AI researchers are more focused on developing systems that are very good at tasks in a narrow range of applications. This focus is at odds with the idea of the pursuit of a super generic AI system that could mimic all different cognitive abilities related to human intelligence such as self-awareness and emotional knowledge. In addition to this debate, about AI development and the status of our hegemony as the most intelligent species on the planet, further societal concerns have been raised. For example, the AI100 (One Hundred Year Study on Artificial Intelligence) a committee led by Stanford University, defined 18 topics of importance for AI [9]. Although these are not exhaustive nor definitive, it sets forth the range of topics that need to be studied, for the potential impact of AI and stresses that there are a number of concerns to be addressed. Many similar assessments have been performed and they each outline similar concerns related to the wider adoption of AI technology.
The 18 topics covered by the AI100
Technical trends and surprises: This topic aims at forecasting the future advances and competencies of AI technologies in the near future. Observatories of the trend and impact of AI should be created, helping to plan the setting of AI in specific sectors, and preparing the necessary regulation to smooth its introduction.
Key opportunities for AI: How advances in AI can help to transform the quality of societal services such as health, education, management and government, covering not just the economic benefits but also the social advantages and impact.
Delays with translating AI advances into real-world values: The pace of translating AI into real world applications is currently driven by potential economic prospects [10]. It is necessary to take measures to foster a rapid translation of those potential applications of AI that can improve or solve a critical need of our society, such as those that can save lives or greatly improve the organisation of social services, even though their economic exploitation is not yet assured.
Privacy and machine intelligence: Personal data and privacy is a major issue to consider and it is important to envisage and prepare the regulatory, legal and policy frameworks related to the sharing of personal data in developing AI systems.
Democracy and freedom: In addition to privacy, ethical questions with respect to the stealth use of AI for unscrupulous applications must be considered. The use of AI should not be at the expense of limiting or influencing the democracy and the freedom of people.
Law: This considers the implications of relevant laws and regulations. First, to identify which aspects of AI require legal assessment and what actions should be undertaken to ensure law enforcement for AI services. It should also provide frameworks and guidelines about how to adhere to the approved laws and policies.
Ethics: By the time AI is deployed into real world applications there are ethical concerns referring to their interaction with the world. What uses of AI should be considered unethical? How should this be disclosed?
Economics: The economic implications of AI on jobs should be monitored and forecasted such that policies can be implemented to direct our future generation into jobs that will not be soon overtaken by machines. The use of sophisticated AI in the financial markets could potentially cause volatilities and it is necessary to assess the influence AI systems may have on financial markets.
AI and warfare: AI has been employed for military applications for more than a decade. Robot snipers and turrets have been developed for military purposes [11]. Intelligent weapons have increasing levels of autonomy and there is a need for developing new conventions and international agreements to define a set of secure boundaries of the use of AI in weaponry and warfare.
Criminal uses of AI: Implementations of AI into malware are becoming more sophisticated thus the chances of stealing personal information from infected devices are getting higher. Malware can be more difficult to detect as evasion techniques by computer viruses and worms may leverage highly sophisticated AI techniques [12, 13]. Another example is the use of drones and their potential to fall into the hands of terrorists the consequence of which would be devastating.
Collaboration with machines: Humans and robots need to work together and it is pertinent to envisage in which scenarios collaboration is critical and how to perform this collaboration safely. Accidents by robots working side by side with people had happened before [14] and robotic and autonomous systems development should focus on not only enhanced task precision but in also being able to understand the environment and human intention.
AI and human cognition: AI has the potential for enhancing human cognitive abilities. Some relevant research disciplines with this objective are sensor informatics and human- computer interfaces. Apart from applications to rehabilitation and assisted living, they are also used in surgery [15] and air traffic control [16]. Cortical implants are increasingly used for controlling prosthesis, our memory and reasoning are increasingly relying on machines and the associated health, safety and ethical impacts must be addressed.
Safety and Autonomy: For the safe operation of intelligent, autonomous systems, formal verification tools should be developed to assess their safety operation. Validation can be focused on the reasoning process and verifying whether the knowledge base of an intelligent system is correct [17] and also making sure that the formulation of the intelligent behaviour will be within safety boundaries [18].
Loss of control of AI systems: The potential of AI being independent from human control is a major concern. Studies should be promoted to address this concern both from the technological standpoint and the relevant framework for governing the responsible development of AI.
Psychology of people and smart machines: More research should be undertaken to obtain detailed knowledge about the opinions and concerns people have, in the wider usage of smart machines in societies. Additionally, in the design of intelligent systems, understanding people’s preferences is important for improving their acceptability [19, 20].
Communication, understanding and outreach: Communication and educational strategies must be developed to embrace AI technologies in our society. These strategies must be formulated in ways that are understandable and accessible by non-experts and the general public.
Neuroscience and AI: Neuroscience and AI can develop together. Neuroscience plays an important role for guiding research in AI and with new advances in high performance computing, there are also new opportunities to study the brain through computational models and simulations in order to investigate new hypotheses [21].
AI and philosophy of mind: When AI can experience a level of consciousness and self-awareness, there will be a need to understand the inner world of the psychology of machines and their subjectivity of consciousness
1. A CLOSER LOOK AT THE EVOLUTION OF AI
SEASONS OF AI
The evolution of AI to date, has endured several cycles of optimism (springs) and pessimism or negativism (winters):
• Birth of AI (1952-1956): Before the term AI was coined, there were already advances in cybernetics and neural networks, which started to attract the attention of both the scientific communities and the public. The Dartmouth Conference (1956) was the result of this increasing interest and gave rise to the following golden years of AI with high levels of optimism in the field.
• First spring (1956-1974): Computers of the time could solve algebra and geometric problems, as well as speak English. Advances were qualified as “impressive” and there was a general atmosphere of optimism in the field. Researchers in the area estimated that a fully intelligent machine would be built in the following 20 years.
• First winter (1974-1980): The winter started when the public and media questioned the promises of AI.
Researchers were caught in a spiral of exaggerated claims and forecasts but the limitations the technology posed at the time were inviolable. An abrupt ending of funding by major agencies such as DARPA, the National Research Council and the British Government, led to the first winter of AI.
• Second Spring (1980-1987): Expert systems were developed to solve problems of a specific domain by using logical rules derived from experts. There was also a revival of connectionism and neural networks for character or speech recognition. This period is known as the second spring of AI.
• Second winter (1987-1993): Specialised machines for running expert systems were displaced by new desktop computers. Consequently some companies, that produced expert systems, went into bankruptcy. This led to a new wave of pessimism ending the funding programs initiated during the previous spring.
• In the background (1997-2000): From 1997 to 2000, the field of AI was progressing behind the scenes, as no further multi-million programs were announced. Despite the lack of major funding the area continued to progress, as increased computer power and resources were developed. New applications in specific areas were developed and the concept of “machine learning” started to become the cornerstone of AI.
• Third spring (2000-Present): Since 2000, with the success of the Internet and web, the Big Data revolution started to take off along with newly emerged areas such as Deep Learning. This new period is known as the third spring of AI and for time being, it looks like it is here to stay. Some have even started to predict the imminent arrival of singularity - an intelligence explosion resulting in a powerful super-intelligence that will eventually surpass human intelligence. Is this possible?
INFLUENCE OF FUNDING
Government organisations and the public sector are investing millions to boost artificial intelligence research. For example, the National Research Foundation of Singapore is investing $150 million into a new national programme in AI. In the UK alone, £270 million is being invested from 2017 to 2018 to boost science, research and innovation, via the Government’s new industrial strategy and a further funding of £4.7 billion is planned by 2021 [22]. This timely investment will put UK in the technological lead among the best in the world and ensure that UK technological innovations can compete. Recent AI developments have triggered major investment across all sectors including financial services, banking, marketing and advertising, in hospitals and government administration.
In fact software and information technology services have more than a 30% share in all AI investments worldwide as of 2016, whereas Internet and telecommunication companies follow with 9% and 4%, respectively [23].
It is also important to note that the funding in AI safety, ethics and strategy/policy has almost doubled in the last three years [24]. Apart from non-profit organisations, such as the Future of Life Institute (FLI) and the Machine Intelligence Research Institute (MIRI), other centres, such as the Centre for Human-Compatible AI and Centre for the Future of Intelligence, have emerged and they, along with key technological firms, invested a total of $6.6 million in 2016.
PUBLICATION VERSUS PATENTING
In terms of international output in publications and patents, there has been a shift of predominant countries influencing the field of AI. In 1995 USA and Japan were the two leading countries in the field of AI patents but this has now shifted to Asia, with China becoming a major player. Since 2010 China and USA have led the way in both scientific publications and in the filing of patents. Other emerging economies, such as India and Brazil, are also rapidly rising.
Recently, there has been a shift relocation of many academic professionals in AI to the industrial sector. The world's largest technology companies have hired several entire research teams previously in universities. These corporations have opted for the publication of pre-prints (ArXiv1 or viXra2) and other non-citable documents, instead of using conventional academic methods of peer-review. This has become an increasing trend. The reason for this is that it allows the prioritisation of claim imprinting without the delay of a peer- review process.
1. FINANCIAL IMPACT OF AI
It has been well recognised that AI amplifies human potential as well as productivity and this is reflected in the rapid increase of investment across many companies and organisations. These include sectors in healthcare,
manufacturing, transport, energy, banking, financial services, management consulting, government administration and marketing/advertising. The revenues of the AI market worldwide, were around 260 billion US dollars in 2016 and this is estimated to exceed $3,060 billion by 2024 [23].
This has had a direct effect on robotic applications, including exoskeletons, rehabilitation, surgical robots and personal care-bots. The economic impact of the next 10 years is estimated to be between $1.49 and $2.95 trillion. These estimates are based on benchmarks that take into account similar technological achievements such as broadband, mobile phones and industrial robots [28]. The investment from the private sector and venture capital is a measure of the market potential of the underlying technology. In 2016, a third of the shares from software and information technology have been invested in AI, whereas in 2015, 1.16 billion US dollars were invested in start-up companies worldwide, a 10-fold increase since 2009.
Major technological firms are investing into applications for speech recognition, natural language processing and computer vision. A significant leap in the performance of machine learning algorithms resulting from deep learning, exploited the improved hardware and sensor technology to train artificial networks with large amounts of information derived from ‘big data’ [31, 32]. Current state-of-the-art AI allows for the automation of various processes and new applications are emerging with the potential to change the entire workings of the business world. As a result, there is huge potential for economic growth, which is demonstrated in the fact that between 2014 and 2015 alone, Google, Microsoft, Apple, Amazon, IBM, Yahoo, Facebook, and Twitter, made at least 26 acquisitions of start-ups and companies developing AI technology, totalling over $5 billion in cost.
In 2014, Google acquired DeepMind, a London-based start-up company specialising in deep learning, for more than $500M and set a record of company investment of AI research to academic standard. In fact, DeepMind has produced over 140 journal and conference papers and has had four articles published in Nature since 2012. One of the achievements of DeepMind was in developing AI technology able to create general-purpose software agents that adjust their actions based only on a cumulative reward. This reinforcement learning approach exceeds human level performance in many aspects and has been demonstrated with the defeat of the world Go game champion; marking a historical landmark in AI progress.
IBM has developed a supercomputer platform, Watson, which has the capability to perform text mining and extract complex analytics from large volumes of unstructured data. To demonstrate its abilities, IBM Watson, in 2011, beat two top players on ‘Jeopardy!’, a popular quiz show, that requires participants to guess questions from specific answers. Although, information retrieval is trivial for computer systems, comprehension of natural language is still a challenge. This achievement has had a significant impact on the performance of web searches and the overall ability of AI systems to interact with humans. In 2015, IBM bought AlchemyAPI to incorporate its text and image analysis capabilities in the cognitive computing platform of the IBM Watson. The system has already been used to process legal documents and provide support to legal duties. Experts believe that these capabilities can transform current health care systems and medical research.
Research in top AI firms is centred on the development of systems that are able to reliably interact with people.
Interaction takes more natural forms through real-time speech recognition and translation capabilities. Robo-advisor applications are at the top of the AI market with a globally estimated 255 billion in US dollars by 2020 [23]. There are already several virtual assistants offered by major companies. For example, Apple offers Siri and Amazon Alexa, Microsoft offers Cortana, and Google has the Google Assistant. In 2016, Apple Inc. purchased Emotient Inc., a start-up using artificial-intelligence technology to read people’s emotions by analyzing facial expressions. DeepMind created WaveNet, which is a generative model that mimics human voices. According to the company’s website, this sounds more natural than the best existing Text-to-Speech systems. Facebook is also considering machine-human interaction capabilities as a prerequisite to generalised AI.
Recently, OpenAI, a non-profit organisation, has been funded as part of a strategic plan to mitigate the risks of monopolising strong AI. OpenAI has re-designed evolutional algorithms that can work together with deep neural networks to offer state-of-the-art performance. It is considered to rival DeepMind since it offers similar open-source machine learning libraries to TensorFlow, a deep learning library distributed by Google DeepMind. Nevertheless, the big difference between the technology developed at OpenAI and the other private tech companies, is that the created Intellectual Property is accessible by everyone.
Although several companies and organisations, including DeepMind and OpenAI, envision the solution to the creation of intelligence and the so-called Strong AI, developing machines with self-sustained long-term goals is well beyond current technology. Furthermore, there is vigorous debate on whether or not we are going through an AI bubble, which encompasses the paradox that productivity growth in USA, during the last decade, has declined regardless of an explosion of technological progress and innovation. It is difficult to understand whether this reflects a statistical shortcoming or that current innovations are not transformative enough. This decline can be also attributed to the lack of consistent policy frameworks and security standards that can enable the application of AI in projects of significant impact.
1. SUBFIELDS AND TECHNOLOGIES THAT UNDERPINNINGS ARTIFICIAL INTELLIGENCE
AI is a diverse field of research and the following subfields are essential to its development. These include neural networks, fuzzy logic, evolutionary computation, and probabilistic methods.
Neural networks build on the area of connectionism with the main purpose of mimicking the way the nervous system processes information. Artificial Neural Networks (ANN) and variants have allowed significant progress of AI to perform tasks relative to “perception”. When combined with the current multicore parallel computing hardware platforms, many neural layers can be stacked to provide a higher level of perceptual abstraction in learning its own set of features, thus removing the need for handcrafted features; a process known as deep learning [33]. Limitations of using deep layered ANN include 1) low interpretability of the resultant learned model, 2) large volumes of training data and considerable computational power are often required for the effective application of these neural models.
Deep learning is part of machine learning and is usually linked to deep neural networks that consist of a multi- level learning of detail or representations of data. Through these different layers, information passes from low-level parameters to higher-level parameters. These different levels correspond to different levels of data abstraction, leading to learning and recognition. A number of deep learning architectures, such as deep neural networks, deep convolutional neural networks and deep belief networks, have been applied to fields such as computer vision, automatic speech recognition, and audio and music signal recognition and these have been shown to produce cutting- edge results in various tasks.
Fuzzy logic focuses on the manipulation of information that is often imprecise. Most computational intelligence principles account for the fact that, whilst observations are always exact, our knowledge of the context, can often be incomplete or inaccurate as it is in many real-world situations. Fuzzy logic provides a framework in which to operate with data assuming a level of imprecision over a set of observations, as well as structural elements to enhance the interpretability of a learned model [34]. It does provide a framework for formalizing AI methods, as well as an accessible translation of AI models into electronic circuits. Nevertheless, fuzzy logic does not provide learning abilities per se, so it is often combined with other aspects such a neural networks, evolutionary computing or statistical learning.
1. THE RISE OF DEEP LEARNING: RETHINKING THE MACHINE LEARNING PIPELINE
The idea of creating an artificial machine is as old as the invention of the computer. Alan Turing in the early 1950s proposed the Turing test, designed to assess whether a machine could be defined as intelligent. Two of the main pioneers in this field are Pitts and McCulloch [38] who, in 1943, developed a technique designed to mimic the way a neuron works. Inspired by this work, a few years later, Frank Rosenblatt [39] developed the first real precursor of the modern neural network, called Perceptron. This algorithm describes an automatic learning procedure that can discriminate linearly separable data. Rosenblatt was confident that the perceptron would lead to an AI system in the future. The introduction of perceptron, in 1958, signalled the beginning of the AI evolution. For almost 10 years afterwards, researchers used this approach to automatically learn how to discriminate data in many applications, until Papert and Minsky [3], demonstrated a few important limitations of Perceptron. This slowed down the fervour of AI progress and more specifically, they proved that the perceptron was not capable of learning simple functions, such as the exclusive-or XOR, no matter how long the network was trained.
Today, we know that the model implied by the perceptron is linear and the XOR function does not belong to this family, but at the time this was enough to stop the research behind neural nets and began the first AI winter. Much later in 1974, the idea of organizing the perceptron in layers and training them using the delta rule [40] shaped the creation of more complex neural nets. With the introduction of the Multilayer Neural Nets [41], researchers were confident that adding multiple hidden layers to the networks would produce deep architectures that further increase the complexity of the hypothesis that can be expressed. However, the hardware constraints that were present at that time, limited, for many years, the number of layers that could be used in practice. To overcome these hardware limitations different network configurations were proposed. For almost another decade, researchers focused on producing new efficient network architectures that are suitable for specific contexts. Notably, these developments included the Autoencoder [42] useful in extracting relevant features from data, the Belief nets used to model statistical variables, the Recurrent neural nets [43] and its variant Long Short Term Memory [44] used for processing sequence of data, and the Convolutional neural nets [45] used to process images. Despite these new AI solutions, the aforementioned hardware limitations were a big restriction during training.
With recent hardware advances, such as the parallelization using GPU, the cloud computing and the multi-core processing finally led to the present stage of AI. In this stage, deep neural nets have made tremendous progress in terms of accuracy and they can now recognize complex images and perform voice translation in real time. However, researchers are still dealing with issues relating to the overfitting of the networks, since large datasets are often required and not always available. Furthermore, with the vanishing of the gradient, this leads to a widespread problem generated during the training of a network with many layers.
For this reason, more sophisticated training procedures have recently been proposed. For example, in 2006, Hinton introduced the idea of unsupervised pretraining and Deep Belief Nets [46]. This approach has each pair of consecutive layers trained separately using an unsupervised model similar to the one used in the Restricted Boltzman Machine [47]; then the obtained parameters are frozen and a new pair of layers are trained and stacked on top of the previous ones. This procedure can be repeated many times leading to the development of a deeper architecture with respect to the traditional neural nets. Moreover, this unsupervised pre-training approach has led to increasing neural net papers when in 2014, for the first time, a neural model became state-of-the-art in the speech recognition.
In 2010, a large database, known as Imagenet containing millions of labelled images was created and this was coupled with an annual challenge called Large Scale Visual Recognition Challenge. This competition requires teams of researchers to build AI systems and they receive a score based on how accurate their model is. In the first two years of the contest, the top models had an error rate of 28% and 26%. In 2012, Krizhevsky, Sutskever and Hinton [48] submitted a solution that had an error rate of just 16% and in 2015 the latest submitted models [49] were capable of beating the human experts with an overall error of 5%. One of the main components of this significant improvement, in such a short time, was due to the extensive use of graphics processing units (GPUs) for speeding up the training procedure, thus allowing the use of larger models which also meant a lower error rate in classification.
In the last 3 years researchers have also been working on training deep neural nets that are capable of beating human experts in different fields, similar to the solution used for AlphaGo [50] or DeepStack [51]and in 2017, they overtook human experts with 44000 played hands of poker.
HIDDEN LAYER
Boltzmann Machines represent a type of neural network modelled by using stochastic units with a specific distribution (for example Gaussian). Learning procedure involves several steps called Gibbs sampling, which gradually adjust the weights to minimize the reconstruction error. They are useful if it is required to model probabilistic relationships between variables. A variant of this machine is the Restricted Boltzmann Machines where the visible and hidden units are restricted to form a bipartite graph that allows implementation of more efficient training algorithms.
An Autoencoder is a neural network designed to extract features directly from the data. This network has the same number of input and output nodes and it is trained using an unsupervised approach to recreate the input vector rather than to assign a class label to it. Usually, the number of hidden units is smaller than the input/output layers, which achieve encoding of the data in a lower dimensional space and extract the most discriminative features.
HIDDEN LAYER
INPUT LAYER
OUTPUT LAYER
CNNs have been proposed to process efficiently imagery data. The name of these networks comes from the convolution operator that provides an easy way to perform complex operations using convolution filter. CNNs use locally connected neurons that represent data specific kernels. The main advantage of a CNN is that during back-propagation, the network has to adjust a number of parameters equal to a single instance of the kernel which drastically reduces the connections from the typical neural network. The concept of CNN is inspired by the neurobiological model of the visual cortex and can be briefly summarized as a sequence of convolution and subsampling of the image until high level features can be extracted.
Recurrent neural nets (RNN)
RNN is a neural network that contains hidden units capable of analysing streams of data. Since RNN suffers from the vanishing gradient and exploding gradient problems, a variation called Long Short-Term Memory units (LSTMs) was proposed in 1997 to solve this problem. Specifically, LSTM is particularly suitable for applications where there are very long time lags of unknown sizes between important events.
RNN and LSTM share the same weights across all steps that greatly reduce the total number of parameters that the network needs to learn. RNNs have shown great successes in many Natural Language Processing tasks such as language modelling, bioinformatics, speech recognition and generating image description.
Ot-2
Ot-1
Ot
Ot+1
OUTPUT STREAM
St-2
St-1
St
St+1
MEMORY
Yt-2
Yt-1
Yt1
Yt+1
INPUT STREAM
1. HARDWARE FOR AI
In 1965, Gordon Moore observed that the number of transistors, in a dense integrated circuit, doubles
approximately every year. Ten years later, he revised his forecast, updating his prediction to the number doubling every two years. Moore's prediction has been accurate for several decades and has been used in the semiconductor industry to guide long-term planning. In 2015, Moore realised that the rate of progress in the hardware would reach saturation and the transistors would arrive at the limits of miniaturisation at the atomic level. Experts estimate that Moore’s law could end in 2025. Today, his prediction is still valid and the number of transistors is increasing even if, after 2005, the frequency and the power started to reduce, leading to a core scaling rather than a frequency improvement. Therefore, since 2005, we are no longer getting faster computers, but the hardware is designed in a multi-core manner. To take full advantage of this different hardware implementation, the software has to be written in a multi-threaded manner too. In future, experts believe that revolutionary technologies may help sustain Moore's law. One of the key challenges will be the design of gates in nanoscale transistors and the ability of controlling the current flow as, when the device dimension shrinks, the connection between transistors becomes more difficult.
1. ROBOTICS AND AI
Building on the advances made in mechatronics, electrical engineering and computing, robotics is developing increasingly sophisticated sensorimotor functions that give machines the ability to adapt to their ever-changing environment. Until now, the system of industrial production was organized around the machine; it is calibrated according to its environment and tolerated minimal variations.
Today, it can be integrated more easily into an existing environment. The autonomy of a robot in an environment can be subdivided into perceiving, planning and execution (manipulating, navigating, collaborating). The main idea of converging AI and Robotics is to try to optimise its level of autonomy through learning. This level of intelligence can be measured as the capacity of predicting the future, either in planning a task, or in interacting (either by manipulating or navigating) with the world. Robots with intelligence have been attempted many times. Although creating a system exhibiting human-like intelligence remains elusive, robots that can perform specialized autonomous tasks, such as driving a vehicle [52], flying in natural and man-made environments [53], swimming [54], carrying boxes and material in different terrains [55], pick up objects [56] and put them down [57]
do exist today.
Another important application of AI in robotics is for the task of perception. Robots can sense the environment by means of integrated sensors or computer vision. In the last decade, computer systems have improved the quality of both sensing and vision. Perception is not only important for planning but also for creating an artificial sense of self-awareness in the robot. This permits supporting interactions with the robot with other entities in the same environment. This discipline is known as social robotics. It covers two broad domains: human-robot interactions (HCI) and cognitive robotics.
The vision of HCI it to improve the robotic perception of humans such as in understanding activities [58], emotions [59], non-verbal communications [60] and in being able to navigate an environment along with humans [61]. The field of cognitive robotics focuses on providing robots with the autonomous capacity of learning and acquiring knowledge from sophisticated levels of perception based on imitation and experience. It aims at mimicking the human cognitive system, which regulates the process of acquiring knowledge and understanding, through experience and sensorisation [62]. In cognitive robotics, there are also models that incorporate motivation and curiosity to improve the quality and speed of knowledge acquisition through learning
[63, 64].
AI has continued beating all records and overcoming many challenges that were unthinkable less than a decade ago. The combination of these advances will continue to reshape our understanding about robotic intelligence in many new domains. Figure 9 provides a timeline of the milestone in robotics and AI.
1. PROGRAMMING LANGUAGES FOR AI
Programming languages played a major role in the evolution of AI since the late 1950s and several teams carried out important research projects in AI; e.g. automatic demonstration programs and game programs (Chess, Ladies) [65]. During these periods researchers found that one of the special requirements for AI is the ability to easily manipulate symbols and lists of symbols rather than processing numbers or strings of characters. Since the languages of the time did not offer such facilities, a researcher from MIT, John MacCarthy, developed, during 1956-58, the definition of an ad-hoc language for logic programming, called LISP (LISt Processing language).
Since then, several hundred derivative languages,
so-called "Lisp dialects", have emerged (Scheme, Common Lisp, Clojure); Indeed, writing a LISP interpreter is not a hard task for a Lisp programmer (it involves only a few thousand instructions) compared to the development of a compiler for a classical language (which requires several tens of thousands of instructions). Because of its expressiveness and flexibility, LISP was very successful in the artificial intelligence community until the 1990s.
Another important event at the beginning of AI was the creation of a language with the main purpose of expressing logic rules and axioms. Around 1972 a new language was created by Alain Colmerauer and Philippe Roussel named
Prolog (PROgramming in Logic). Their goal was to create a programming language where the expected logical rules of a solution can be defined and the compiler automatically transforms it into a sequence of instructions. Prolog is used in AI and in natural language processing. Its rules of syntax and its semantics are simple and considered accessible to non-programmers. One of the objectives was to provide a tool for linguistics that was compatible with computer science.
In the 1990s, the machine languages with C / C ++ and Fortran gained popularity and eclipsed the use of LISP and Prolog. Greater emphasis was placed on creating functions and libraries for scientific computation on these platforms and were used for intensive data analysis tasks or artificial intelligence with early robots. In the middle of the 1990s, the company Sun Microsystems, started a project to create a language that solved secutiry flaws, distributed programming and multi-threading of C++. In addition, they wanted a platform that could be ported to any type of device or platform. In 1995, they presented Java, which took the concept of object orientation much further than C++.
Equally, one of the most important additions to Java was the Java VM (JVM) which enabled the capability of running the same code in any device regardless of their internal technology and without the need of pre-compiling for every platform. This added new advantages to the field of AI that were be introduced in devices such as cloud servers and embedded computers. Another important feature of Java was that it also offered one of the first frameworks, with specific tools for the internet, bringing the possibility of running applications in the form of java applets and javascripts (i.e. self-executing programs) without the need of installation. This had an enormous impact in the field of AI and a set the foundation in the fields of web 2.0/3.0 and the internet of things (IoT).
However, the development of AI using purely procedural languages was costly, time-consuming and error prone. Consequently, this turned the attention into other multi- paradigm languages that could combine features from functional and procedural object-oriented languages.
Python, although first published in 1991, started to gain popularity as an alternative to C/C++ with Python 2.2 by 2001. The Python concept was to have a language that could be as powerful as C/C++ but also expressive and pragmatic for executing "scripts" like Shell Script. It was in 2008, with the publication of Python 3.0, which solved several initial flaws, when the language started to be considered a serious contender for C++, java and other scripting languages such as Perl.
Since 2008, the Python community has been trying to catch up with specific languages for scientific computing, such as Matlab and R. Due to its versatility, Python is now used frequently for research in AI. However, although python has some of the advantages of functional programming, run-time speeds are still far behind other functional languages, such as Lisp or Haskell, and even more so from C/C++.
In addition, it lacks of efficiency when managing large amounts of memory and highly-concurrent systems.
From 2010 and mostly driven by the necessity of translating AI into commercial products, (that could be used by thousands and millions of users in real time), IT corporations looked for alternatives by creating hybrid languages, that combined the best from all paradigms without compromising speed, capacity and concurrency. In recent years, new languages such as Scala and Go, as well as Erlang or Clojure, have been used for applications with very high concurrency and parallelization, mostly on the server side.
Well-known examples are Facebook with Erlang or Google with Go. New languages for scientific computation have also emerged such as Julia and Lua.
Although functional programming has been popular in academia, its use in industrial settings has been marginal and mainly during the times when “expert systems” were at their peak, predominantly during the 1980s. After the fall of expert systems, functional programing has, for many years, been considered a failing relic from that period. However, as multiprocessors and parallel computing are becoming more available, functional programming is proving to be a choice of many programmers to maximise functionality from their multicore processors. These highly expensive computations are usually needed for heavy mathematical operations or pattern matching, which constitute a fundamental part of running an AI system. In the future, we will see new languages that bring simplifications on existing functional languages such as Haskell and Erlang and make this programming paradigm more accessible. In addition, the advent of the internet-of-things (IoT) has drawn the attention to the programming of embedded systems. Thus, efficiency, safety and performance are again matters for discussion.
New languages that can replace C/C++ incorporating tips from functional programming (e.g. Elixir) will become
increasingly popular. Also, new languages that incorporate simplifications as well as a set of functions from modern imperative programming, while maintaining a performance like C/C++ (e.g. Rust), will be another future development.
1. IMPACT OF MACHINE VISION
Machine vision integrates image capture systems with computer vision algorithms to provide automatic inspection and robot guidance. Although it is inspired by the human vision system, based on the extraction of conceptual information from two-dimensional images, machine vision systems are not restricted to 2D visible light. Optical sensors include single beam lasers to 3D high definition Light Detection And Ranging (LiDAR) systems, also known as laser scanning 2D or 3D sonar sensors and one or multiple 2D camera systems. Nevertheless, most machine vision applications are based on 2D image-based capture systems and computer vision algorithms that mimic aspects of human visual perception. Humans perceive the surrounding world in 3D and their ability to navigate and accomplish certain tasks depends on reconstructing 3D information from 2D images that allows them to locate themselves in relation to the surrounding objects. Subsequently, this information is combined with prior knowledge in order to detect and identify objects around them and understand how they interact. Scene reconstruction along with object detection and recognition are the main sub-domains of computer vision.
Regardless of the imaging sensors used, the most common approaches of reconstructing 3D information are normally based on either time-of-flight techniques, multi-view geometry and/or on photometric stereo. The former is used in laser scanners to estimate the distance between the light source and the object based on the time required for the light to reach the object and return back. Time-of-flight approaches are used to measure distances in kilometres and they are accurate to a millimetre scale, since they are limited by the ability to measure time. On the other hand, multi-view geometry problems include ‘structure’ problems, ‘stereo correspondence’ problems and ‘motion’ problems. Recovery of the 3D ‘structure’ implies that given 2D projections of the same 3D point, in two or more images, the 3D coordinates of the point are estimated based on triangulation. ‘Stereo correspondence’ refers to the problem of finding the image point that corresponds to a point from another 2D view.
Finally, ‘motion’ refers to the problem of recovering the camera coordinates given a set of corresponding points in two or more image views. 3D laser scanners based on triangulation can reach micrometre accuracy but their range is constrained to a few meters. Several sub-problems such as ‘structure from motion’ uses multi-view geometry principles to extract corresponding points between 2D views of the same object and reconstruct its shape.
Stereo-vision assumes the robust extraction of corresponding salient points/features across images, the so-called interest point detection. These features should be invariant to photometric transformation such as changes in the lighting conditions and covariant to geometric transformations.
For over two decades researchers have proposed several approaches. The Scale-invariant feature transform (SIFT) extracts features that are invariant to scale, rotation and translation transformations and robust to illumination variations and moderate perspective transformations.
Since its introduction in 1999-2004, it has been successful in several vision applications, including object recognition, robot localisation and mapping.
Representing and recognising object categories have proven much harder problems to generalise and solve, compared with 3D reconstruction, since there are thousands of objects that can belong to an arbitrary number of categories simultaneously. Several ideas about object detection are related to Gestalt psychology, which is a theory of mind
with relation to visual perception. A major aspect of the theory is about grouping entities together based on their proximity, similarity, symmetry, common fate, continuity and so on. From the 1960s to early 1990s, research in object recognition was centred on geometric shapes. This was a bottom-up process, which uses a small number of primitive 3D dimensional objects that are assembled together in various configurations to form complex objects. In the 1990s, appearance-based models were explored, which were based on manifold learning of the object appearance parameterised by the pose and illumination [66]. These techniques are not robust to occlusion, clutter and deformation. By the mid-late 1990s, sliding window approaches were designed that classify whether an object is found for each instance of a sliding window across an image [67]. The main challenges were how to design features that represent appropriately the appearance of the object and how to efficiently search a large number of positions and scales. Local features approaches were also developed and they aimed towards those which were invariant to image scaling, geometric transformations and illumination changes [68]. In the early 2000s 'parts-and- shape' models along with 'bags of features' were suggested. Parts-and-shape models represent complex objects using combinations of multi-scaled deformable objects [69].
On the other hand, bags of features methods, represent visual features as words and relate object recognition and image classification to the expressive power of natural language processing approaches [70].
Machine learning in object recognition facilitated a shift, from solving a problem based on mathematical modelling alone, to learning algorithms based on real-data and statistical modelling. A major breakthrough in object recognition and classification came in 2012 with the emergence of deep neural networks and the availability of large labelled image databases, such as ImageNet. Compared to classical object recognition methods, which depend on feature extraction followed by feature matching methodologies, deep learning has the advantage of encoding both feature extraction and image classification via the structure of a neural network.
The superb performance of deep neural networks resulted in an increase of image classification from 72% in 2010 to 96% in 2015, which outperforms human accuracy and has had a significant impact in real-life applications [71]. Both Google and Baidu updated their image search capabilities based on the Hinton’s deep neural network architecture. Face detection has been introduced in several mobile devices and Apple even created an app to recognise pets. The accuracy of these models in object recognition and image classification exceed human-level accuracy and spread waves of technological changes across the industry.
The term AI was closely associated with the field of “symbolic AI”, which was popular until the end of the 1980s. In order to overcome some of the limitations of symbolic AI, subsymbolic methodologies such as neural networks, fuzzy systems, evolutionary computation and other computational models started gaining popularity, leading to the term “computational intelligence” emerging as a subfield of AI.
Nowadays, the term AI encompasses the whole conceptualisation of a machine that is intelligent in terms of both operational and social consequences. A practical definition used is one proposed by Russell and Norvig: “Artificial Intelligence is the study of human intelligence and actions replicated artificially, such that the resultant bears to its design a reasonable level of rationality” [1]. This definition can be further refined by stipulating that the level of rationality may even supersede humans, for specific and well-defined tasks.
Current AI technologies are used in online advertising, driving, aviation, medicine and personal assistance image recognition. The recent success of AI has captured the imagination of both the scientific community and the public. An example of this is vehicles equipped with an automatic steering system, also known as autonomous cars. Each vehicle is equipped with a series of lidar sensors and cameras which enable recognition of its three-dimensional environment and provides the ability to make intelligent decisions on maneuvers in variable, real-traffic road conditions. Another example is the Alpha-Go, developed by Google Deepmind, to play the board game Go. Last year, Alpha-Go defeated the Korean grandmaster Lee Sedol, becoming the first machine to beat a professional player and recently it went on to win against the current world number one, Ke Jie, in China. The number of possible games in Go is estimated to be 10761 and given the extreme complexity of the game, most AI researchers believed it would be years before this could happen. This has led to both the excitement and fear in many that AI will surpass humans in all the fields it marches into.
However, current AI technologies are limited to very specific applications. One limitation of AI, for example, is the lack of “common sense”; the ability to judge information beyond its acquired knowledge. A recent example is that of the AI robot Tay developed by Microsoft and designed for making conversations on social networks. It had to be disconnected shortly after its launch because it was not able to distinguish between positive and negative human interaction. AI is also limited in terms of emotional intelligence. AI can only detect basic human emotional states such as anger, joy, sadness, fear, pain, stress and neutrality. Emotional intelligence is one of the next frontiers of higher levels of personalisation.
True and complete AI does not yet exist. At this level, AI will mimic human cognition to a point that it will enable the ability to dream, think, feel emotions and have own goals. Although there is no evidence yet this kind of true AI could exist before 2050, nevertheless the computer science principles driving AI forward, are rapidly advancing and it is important to assess its impact, not only from a technological standpoint, but also from a social, ethical and legal perspective.
THE BIRTH AND BOOM OF AI
The birth of the computer took place when the first calculator machines were developed, from the mechanical calculator of Babbage, to the electromechanical calculator of Torres-Quevedo. The dawn of automata theory can be traced back to World War II with what was known as the “codebreakers”. The amount of operations required to decode the German trigrams of the Enigma machine, without knowing the rotor’s position, proved to be too challenging to be solved manually. The inclusion of automata theory in computing conceived the first logical machines to account for operations such as generating, codifying, storing and using information. Indeed, these four tasks are the basic operations of information processing performed by humans. The pioneering work by Ramón y Cajal marked the birth of neuroscience, although many neurological structures and stimulus responses were already known and studied before him. For the first time in history the concept of “neuron” was proposed. McClulloch and Pitts further developed a connection between automata theory and neuroscience, proposing the first artificial neuron which, years later, gave rise to the first computational intelligence algorithm, namely “the perceptron”. This idea generated great interest among prominent scientists of the time, such as Von Neumann, who was the pioneer of modern computers and set the foundation for the connectionism movement.
In 1971, DARPA funded a consortium of leading laboratories in the field of speech recognition. The project had the ambitious goal of creating a fully functional speech recognition system with a large vocabulary. In the middle of the 1970s, the field of AI endured fierce criticism and budgetary restrictions, as AI research development did not match the overwhelming expectations of researchers.. When promised results did not materialize, investment in AI eroded. Following disappointing results, DARPA withdrew funding in speech recognition and this, coupled with other events such as the failure of machine translation, the abandonment of connectionism and the Lighthill report, marked the first winter of AI [2]. During this period, connectionism stagnated for the next 10 years following a devastating critique by Marvin Minksy on perceptrons [3].
From 1980 until 1987, AI programmes, called “expert systems”, were adopted by companies and knowledge acquisition become the central focus of AI research. At the same time, the Japanese government launched a massive funding program on AI, with its fifth-generation computers initiative. Connectionism was also revived by the work of John Hopfield [4] and David Rumelhart [5].
AI researchers who had experienced the first backlash in collapsed and billions of dollars were lost. The difficulty of updating and reprograming the expert systems, in addition to the high maintenance costs, led to the second AI winter. Investment in AI dropped and DARPA stopped its strategic computing initiative, claiming AI was no longer the “latest mode”. Japan also stopped funding its fifth-generation computer program as the proposed goals were not achieved.
In the 1990s, the new concept of “intelligent agent” emerged [6]. An agent is a system that perceives its environment
and undertakes actions that maximize its chances of being successful. The concept of agents conveys, for the first time, the idea of intelligent units working collaboratively with a common objective. This new paradigm was intended to mimic how humans work collectively in groups, organizations and/or societies. Intelligent agents proved to be a more polyvalent concept of intelligence. In the late 1990s, fields such as statistical learning from several perspectives including probabilistic, frequentist and possibilistic (fuzzy logic) approaches, were linked to AI to deal with the uncertainty of decisions. This brought a new wave of successful applications for AI, beyond what expert systems had achieved during the 1980s. These new ways of reasoning were more suited to cope with the uncertainty of intelligent agent states and perceptions and had its major impact in the field of control. During this time, high-speed trains controlled by fuzzy logic, were developed [7] as were many other industrial applications (e.g. factory valves, gas and petrol tanks surveillance, automatic gear transmission systems and reactor control in power plants) as well as household appliances with advanced levels of intelligence (e.g. air-conditioners, heating systems, cookers and vacuum- cleaners). These were different to the expert systems in 1980s; the modelling of the inference system for the task, achieved through learning, gave rise to the field of Machine Learning. Nevertheless, although machine reasoning exhibited good performance, there was still an engineering requirement to digest the input space into a new source, so that intelligence could reason more effectively. Since 2000, a third renaissance of the connectionism paradigm arrived with the dawn of Big Data, propelled by the rapid adoption of the Internet and mobile communication. Neural networks were once more considered, particularly in the role they played in enhancing perceptual intelligence and eliminating the necessity of feature engineering. Great advances were also made in computer vision, improving visual perception, increasing the capabilities of intelligent agents and robots in performing more complex tasks, combined with visual pattern recognition. All these paved the way to new AI challenges such as, speech recognition, natural language processing, and self-driving cars. A timeline of key highlights in the history of AI is shown in Figure 1.
1. QUESTIONING THE IMPACT OF AI
Given the exponential rise of interest in AI, experts have called for major studies on the impact of AI on our society, not only in technological but also in legal, ethical and socio- economic areas. This response also includes the speculation that autonomous super artificial intelligence may one day supersede the cognitive capabilities of humans. This future scenario is usually known in AI forums as the “AI singularity” [8]. This is commonly defined as the ability of machines to build better machines by themselves. This futuristic scenario has been questioned and is received with scepticism by many experts. Today’s AI researchers are more focused on developing systems that are very good at tasks in a narrow range of applications. This focus is at odds with the idea of the pursuit of a super generic AI system that could mimic all different cognitive abilities related to human intelligence such as self-awareness and emotional knowledge. In addition to this debate, about AI development and the status of our hegemony as the most intelligent species on the planet, further societal concerns have been raised. For example, the AI100 (One Hundred Year Study on Artificial Intelligence) a committee led by Stanford University, defined 18 topics of importance for AI [9]. Although these are not exhaustive nor definitive, it sets forth the range of topics that need to be studied, for the potential impact of AI and stresses that there are a number of concerns to be addressed. Many similar assessments have been performed and they each outline similar concerns related to the wider adoption of AI technology.
The 18 topics covered by the AI100
Technical trends and surprises: This topic aims at forecasting the future advances and competencies of AI technologies in the near future. Observatories of the trend and impact of AI should be created, helping to plan the setting of AI in specific sectors, and preparing the necessary regulation to smooth its introduction.
Key opportunities for AI: How advances in AI can help to transform the quality of societal services such as health, education, management and government, covering not just the economic benefits but also the social advantages and impact.
Delays with translating AI advances into real-world values: The pace of translating AI into real world applications is currently driven by potential economic prospects [10]. It is necessary to take measures to foster a rapid translation of those potential applications of AI that can improve or solve a critical need of our society, such as those that can save lives or greatly improve the organisation of social services, even though their economic exploitation is not yet assured.
Privacy and machine intelligence: Personal data and privacy is a major issue to consider and it is important to envisage and prepare the regulatory, legal and policy frameworks related to the sharing of personal data in developing AI systems.
Democracy and freedom: In addition to privacy, ethical questions with respect to the stealth use of AI for unscrupulous applications must be considered. The use of AI should not be at the expense of limiting or influencing the democracy and the freedom of people.
Law: This considers the implications of relevant laws and regulations. First, to identify which aspects of AI require legal assessment and what actions should be undertaken to ensure law enforcement for AI services. It should also provide frameworks and guidelines about how to adhere to the approved laws and policies.
Ethics: By the time AI is deployed into real world applications there are ethical concerns referring to their interaction with the world. What uses of AI should be considered unethical? How should this be disclosed?
Economics: The economic implications of AI on jobs should be monitored and forecasted such that policies can be implemented to direct our future generation into jobs that will not be soon overtaken by machines. The use of sophisticated AI in the financial markets could potentially cause volatilities and it is necessary to assess the influence AI systems may have on financial markets.
AI and warfare: AI has been employed for military applications for more than a decade. Robot snipers and turrets have been developed for military purposes [11]. Intelligent weapons have increasing levels of autonomy and there is a need for developing new conventions and international agreements to define a set of secure boundaries of the use of AI in weaponry and warfare.
Criminal uses of AI: Implementations of AI into malware are becoming more sophisticated thus the chances of stealing personal information from infected devices are getting higher. Malware can be more difficult to detect as evasion techniques by computer viruses and worms may leverage highly sophisticated AI techniques [12, 13]. Another example is the use of drones and their potential to fall into the hands of terrorists the consequence of which would be devastating.
Collaboration with machines: Humans and robots need to work together and it is pertinent to envisage in which scenarios collaboration is critical and how to perform this collaboration safely. Accidents by robots working side by side with people had happened before [14] and robotic and autonomous systems development should focus on not only enhanced task precision but in also being able to understand the environment and human intention.
AI and human cognition: AI has the potential for enhancing human cognitive abilities. Some relevant research disciplines with this objective are sensor informatics and human- computer interfaces. Apart from applications to rehabilitation and assisted living, they are also used in surgery [15] and air traffic control [16]. Cortical implants are increasingly used for controlling prosthesis, our memory and reasoning are increasingly relying on machines and the associated health, safety and ethical impacts must be addressed.
Safety and Autonomy: For the safe operation of intelligent, autonomous systems, formal verification tools should be developed to assess their safety operation. Validation can be focused on the reasoning process and verifying whether the knowledge base of an intelligent system is correct [17] and also making sure that the formulation of the intelligent behaviour will be within safety boundaries [18].
Loss of control of AI systems: The potential of AI being independent from human control is a major concern. Studies should be promoted to address this concern both from the technological standpoint and the relevant framework for governing the responsible development of AI.
Psychology of people and smart machines: More research should be undertaken to obtain detailed knowledge about the opinions and concerns people have, in the wider usage of smart machines in societies. Additionally, in the design of intelligent systems, understanding people’s preferences is important for improving their acceptability [19, 20].
Communication, understanding and outreach: Communication and educational strategies must be developed to embrace AI technologies in our society. These strategies must be formulated in ways that are understandable and accessible by non-experts and the general public.
Neuroscience and AI: Neuroscience and AI can develop together. Neuroscience plays an important role for guiding research in AI and with new advances in high performance computing, there are also new opportunities to study the brain through computational models and simulations in order to investigate new hypotheses [21].
AI and philosophy of mind: When AI can experience a level of consciousness and self-awareness, there will be a need to understand the inner world of the psychology of machines and their subjectivity of consciousness
1. A CLOSER LOOK AT THE EVOLUTION OF AI
SEASONS OF AI
The evolution of AI to date, has endured several cycles of optimism (springs) and pessimism or negativism (winters):
• Birth of AI (1952-1956): Before the term AI was coined, there were already advances in cybernetics and neural networks, which started to attract the attention of both the scientific communities and the public. The Dartmouth Conference (1956) was the result of this increasing interest and gave rise to the following golden years of AI with high levels of optimism in the field.
• First spring (1956-1974): Computers of the time could solve algebra and geometric problems, as well as speak English. Advances were qualified as “impressive” and there was a general atmosphere of optimism in the field. Researchers in the area estimated that a fully intelligent machine would be built in the following 20 years.
• First winter (1974-1980): The winter started when the public and media questioned the promises of AI.
Researchers were caught in a spiral of exaggerated claims and forecasts but the limitations the technology posed at the time were inviolable. An abrupt ending of funding by major agencies such as DARPA, the National Research Council and the British Government, led to the first winter of AI.
• Second Spring (1980-1987): Expert systems were developed to solve problems of a specific domain by using logical rules derived from experts. There was also a revival of connectionism and neural networks for character or speech recognition. This period is known as the second spring of AI.
• Second winter (1987-1993): Specialised machines for running expert systems were displaced by new desktop computers. Consequently some companies, that produced expert systems, went into bankruptcy. This led to a new wave of pessimism ending the funding programs initiated during the previous spring.
• In the background (1997-2000): From 1997 to 2000, the field of AI was progressing behind the scenes, as no further multi-million programs were announced. Despite the lack of major funding the area continued to progress, as increased computer power and resources were developed. New applications in specific areas were developed and the concept of “machine learning” started to become the cornerstone of AI.
• Third spring (2000-Present): Since 2000, with the success of the Internet and web, the Big Data revolution started to take off along with newly emerged areas such as Deep Learning. This new period is known as the third spring of AI and for time being, it looks like it is here to stay. Some have even started to predict the imminent arrival of singularity - an intelligence explosion resulting in a powerful super-intelligence that will eventually surpass human intelligence. Is this possible?
INFLUENCE OF FUNDING
Government organisations and the public sector are investing millions to boost artificial intelligence research. For example, the National Research Foundation of Singapore is investing $150 million into a new national programme in AI. In the UK alone, £270 million is being invested from 2017 to 2018 to boost science, research and innovation, via the Government’s new industrial strategy and a further funding of £4.7 billion is planned by 2021 [22]. This timely investment will put UK in the technological lead among the best in the world and ensure that UK technological innovations can compete. Recent AI developments have triggered major investment across all sectors including financial services, banking, marketing and advertising, in hospitals and government administration.
In fact software and information technology services have more than a 30% share in all AI investments worldwide as of 2016, whereas Internet and telecommunication companies follow with 9% and 4%, respectively [23].
It is also important to note that the funding in AI safety, ethics and strategy/policy has almost doubled in the last three years [24]. Apart from non-profit organisations, such as the Future of Life Institute (FLI) and the Machine Intelligence Research Institute (MIRI), other centres, such as the Centre for Human-Compatible AI and Centre for the Future of Intelligence, have emerged and they, along with key technological firms, invested a total of $6.6 million in 2016.
PUBLICATION VERSUS PATENTING
In terms of international output in publications and patents, there has been a shift of predominant countries influencing the field of AI. In 1995 USA and Japan were the two leading countries in the field of AI patents but this has now shifted to Asia, with China becoming a major player. Since 2010 China and USA have led the way in both scientific publications and in the filing of patents. Other emerging economies, such as India and Brazil, are also rapidly rising.
Recently, there has been a shift relocation of many academic professionals in AI to the industrial sector. The world's largest technology companies have hired several entire research teams previously in universities. These corporations have opted for the publication of pre-prints (ArXiv1 or viXra2) and other non-citable documents, instead of using conventional academic methods of peer-review. This has become an increasing trend. The reason for this is that it allows the prioritisation of claim imprinting without the delay of a peer- review process.
1. FINANCIAL IMPACT OF AI
It has been well recognised that AI amplifies human potential as well as productivity and this is reflected in the rapid increase of investment across many companies and organisations. These include sectors in healthcare,
manufacturing, transport, energy, banking, financial services, management consulting, government administration and marketing/advertising. The revenues of the AI market worldwide, were around 260 billion US dollars in 2016 and this is estimated to exceed $3,060 billion by 2024 [23].
This has had a direct effect on robotic applications, including exoskeletons, rehabilitation, surgical robots and personal care-bots. The economic impact of the next 10 years is estimated to be between $1.49 and $2.95 trillion. These estimates are based on benchmarks that take into account similar technological achievements such as broadband, mobile phones and industrial robots [28]. The investment from the private sector and venture capital is a measure of the market potential of the underlying technology. In 2016, a third of the shares from software and information technology have been invested in AI, whereas in 2015, 1.16 billion US dollars were invested in start-up companies worldwide, a 10-fold increase since 2009.
Major technological firms are investing into applications for speech recognition, natural language processing and computer vision. A significant leap in the performance of machine learning algorithms resulting from deep learning, exploited the improved hardware and sensor technology to train artificial networks with large amounts of information derived from ‘big data’ [31, 32]. Current state-of-the-art AI allows for the automation of various processes and new applications are emerging with the potential to change the entire workings of the business world. As a result, there is huge potential for economic growth, which is demonstrated in the fact that between 2014 and 2015 alone, Google, Microsoft, Apple, Amazon, IBM, Yahoo, Facebook, and Twitter, made at least 26 acquisitions of start-ups and companies developing AI technology, totalling over $5 billion in cost.
In 2014, Google acquired DeepMind, a London-based start-up company specialising in deep learning, for more than $500M and set a record of company investment of AI research to academic standard. In fact, DeepMind has produced over 140 journal and conference papers and has had four articles published in Nature since 2012. One of the achievements of DeepMind was in developing AI technology able to create general-purpose software agents that adjust their actions based only on a cumulative reward. This reinforcement learning approach exceeds human level performance in many aspects and has been demonstrated with the defeat of the world Go game champion; marking a historical landmark in AI progress.
IBM has developed a supercomputer platform, Watson, which has the capability to perform text mining and extract complex analytics from large volumes of unstructured data. To demonstrate its abilities, IBM Watson, in 2011, beat two top players on ‘Jeopardy!’, a popular quiz show, that requires participants to guess questions from specific answers. Although, information retrieval is trivial for computer systems, comprehension of natural language is still a challenge. This achievement has had a significant impact on the performance of web searches and the overall ability of AI systems to interact with humans. In 2015, IBM bought AlchemyAPI to incorporate its text and image analysis capabilities in the cognitive computing platform of the IBM Watson. The system has already been used to process legal documents and provide support to legal duties. Experts believe that these capabilities can transform current health care systems and medical research.
Research in top AI firms is centred on the development of systems that are able to reliably interact with people.
Interaction takes more natural forms through real-time speech recognition and translation capabilities. Robo-advisor applications are at the top of the AI market with a globally estimated 255 billion in US dollars by 2020 [23]. There are already several virtual assistants offered by major companies. For example, Apple offers Siri and Amazon Alexa, Microsoft offers Cortana, and Google has the Google Assistant. In 2016, Apple Inc. purchased Emotient Inc., a start-up using artificial-intelligence technology to read people’s emotions by analyzing facial expressions. DeepMind created WaveNet, which is a generative model that mimics human voices. According to the company’s website, this sounds more natural than the best existing Text-to-Speech systems. Facebook is also considering machine-human interaction capabilities as a prerequisite to generalised AI.
Recently, OpenAI, a non-profit organisation, has been funded as part of a strategic plan to mitigate the risks of monopolising strong AI. OpenAI has re-designed evolutional algorithms that can work together with deep neural networks to offer state-of-the-art performance. It is considered to rival DeepMind since it offers similar open-source machine learning libraries to TensorFlow, a deep learning library distributed by Google DeepMind. Nevertheless, the big difference between the technology developed at OpenAI and the other private tech companies, is that the created Intellectual Property is accessible by everyone.
Although several companies and organisations, including DeepMind and OpenAI, envision the solution to the creation of intelligence and the so-called Strong AI, developing machines with self-sustained long-term goals is well beyond current technology. Furthermore, there is vigorous debate on whether or not we are going through an AI bubble, which encompasses the paradox that productivity growth in USA, during the last decade, has declined regardless of an explosion of technological progress and innovation. It is difficult to understand whether this reflects a statistical shortcoming or that current innovations are not transformative enough. This decline can be also attributed to the lack of consistent policy frameworks and security standards that can enable the application of AI in projects of significant impact.
1. SUBFIELDS AND TECHNOLOGIES THAT UNDERPINNINGS ARTIFICIAL INTELLIGENCE
AI is a diverse field of research and the following subfields are essential to its development. These include neural networks, fuzzy logic, evolutionary computation, and probabilistic methods.
Neural networks build on the area of connectionism with the main purpose of mimicking the way the nervous system processes information. Artificial Neural Networks (ANN) and variants have allowed significant progress of AI to perform tasks relative to “perception”. When combined with the current multicore parallel computing hardware platforms, many neural layers can be stacked to provide a higher level of perceptual abstraction in learning its own set of features, thus removing the need for handcrafted features; a process known as deep learning [33]. Limitations of using deep layered ANN include 1) low interpretability of the resultant learned model, 2) large volumes of training data and considerable computational power are often required for the effective application of these neural models.
Deep learning is part of machine learning and is usually linked to deep neural networks that consist of a multi- level learning of detail or representations of data. Through these different layers, information passes from low-level parameters to higher-level parameters. These different levels correspond to different levels of data abstraction, leading to learning and recognition. A number of deep learning architectures, such as deep neural networks, deep convolutional neural networks and deep belief networks, have been applied to fields such as computer vision, automatic speech recognition, and audio and music signal recognition and these have been shown to produce cutting- edge results in various tasks.
Fuzzy logic focuses on the manipulation of information that is often imprecise. Most computational intelligence principles account for the fact that, whilst observations are always exact, our knowledge of the context, can often be incomplete or inaccurate as it is in many real-world situations. Fuzzy logic provides a framework in which to operate with data assuming a level of imprecision over a set of observations, as well as structural elements to enhance the interpretability of a learned model [34]. It does provide a framework for formalizing AI methods, as well as an accessible translation of AI models into electronic circuits. Nevertheless, fuzzy logic does not provide learning abilities per se, so it is often combined with other aspects such a neural networks, evolutionary computing or statistical learning.
1. THE RISE OF DEEP LEARNING: RETHINKING THE MACHINE LEARNING PIPELINE
The idea of creating an artificial machine is as old as the invention of the computer. Alan Turing in the early 1950s proposed the Turing test, designed to assess whether a machine could be defined as intelligent. Two of the main pioneers in this field are Pitts and McCulloch [38] who, in 1943, developed a technique designed to mimic the way a neuron works. Inspired by this work, a few years later, Frank Rosenblatt [39] developed the first real precursor of the modern neural network, called Perceptron. This algorithm describes an automatic learning procedure that can discriminate linearly separable data. Rosenblatt was confident that the perceptron would lead to an AI system in the future. The introduction of perceptron, in 1958, signalled the beginning of the AI evolution. For almost 10 years afterwards, researchers used this approach to automatically learn how to discriminate data in many applications, until Papert and Minsky [3], demonstrated a few important limitations of Perceptron. This slowed down the fervour of AI progress and more specifically, they proved that the perceptron was not capable of learning simple functions, such as the exclusive-or XOR, no matter how long the network was trained.
Today, we know that the model implied by the perceptron is linear and the XOR function does not belong to this family, but at the time this was enough to stop the research behind neural nets and began the first AI winter. Much later in 1974, the idea of organizing the perceptron in layers and training them using the delta rule [40] shaped the creation of more complex neural nets. With the introduction of the Multilayer Neural Nets [41], researchers were confident that adding multiple hidden layers to the networks would produce deep architectures that further increase the complexity of the hypothesis that can be expressed. However, the hardware constraints that were present at that time, limited, for many years, the number of layers that could be used in practice. To overcome these hardware limitations different network configurations were proposed. For almost another decade, researchers focused on producing new efficient network architectures that are suitable for specific contexts. Notably, these developments included the Autoencoder [42] useful in extracting relevant features from data, the Belief nets used to model statistical variables, the Recurrent neural nets [43] and its variant Long Short Term Memory [44] used for processing sequence of data, and the Convolutional neural nets [45] used to process images. Despite these new AI solutions, the aforementioned hardware limitations were a big restriction during training.
With recent hardware advances, such as the parallelization using GPU, the cloud computing and the multi-core processing finally led to the present stage of AI. In this stage, deep neural nets have made tremendous progress in terms of accuracy and they can now recognize complex images and perform voice translation in real time. However, researchers are still dealing with issues relating to the overfitting of the networks, since large datasets are often required and not always available. Furthermore, with the vanishing of the gradient, this leads to a widespread problem generated during the training of a network with many layers.
For this reason, more sophisticated training procedures have recently been proposed. For example, in 2006, Hinton introduced the idea of unsupervised pretraining and Deep Belief Nets [46]. This approach has each pair of consecutive layers trained separately using an unsupervised model similar to the one used in the Restricted Boltzman Machine [47]; then the obtained parameters are frozen and a new pair of layers are trained and stacked on top of the previous ones. This procedure can be repeated many times leading to the development of a deeper architecture with respect to the traditional neural nets. Moreover, this unsupervised pre-training approach has led to increasing neural net papers when in 2014, for the first time, a neural model became state-of-the-art in the speech recognition.
In 2010, a large database, known as Imagenet containing millions of labelled images was created and this was coupled with an annual challenge called Large Scale Visual Recognition Challenge. This competition requires teams of researchers to build AI systems and they receive a score based on how accurate their model is. In the first two years of the contest, the top models had an error rate of 28% and 26%. In 2012, Krizhevsky, Sutskever and Hinton [48] submitted a solution that had an error rate of just 16% and in 2015 the latest submitted models [49] were capable of beating the human experts with an overall error of 5%. One of the main components of this significant improvement, in such a short time, was due to the extensive use of graphics processing units (GPUs) for speeding up the training procedure, thus allowing the use of larger models which also meant a lower error rate in classification.
In the last 3 years researchers have also been working on training deep neural nets that are capable of beating human experts in different fields, similar to the solution used for AlphaGo [50] or DeepStack [51]and in 2017, they overtook human experts with 44000 played hands of poker.
HIDDEN LAYER
Boltzmann Machines represent a type of neural network modelled by using stochastic units with a specific distribution (for example Gaussian). Learning procedure involves several steps called Gibbs sampling, which gradually adjust the weights to minimize the reconstruction error. They are useful if it is required to model probabilistic relationships between variables. A variant of this machine is the Restricted Boltzmann Machines where the visible and hidden units are restricted to form a bipartite graph that allows implementation of more efficient training algorithms.
An Autoencoder is a neural network designed to extract features directly from the data. This network has the same number of input and output nodes and it is trained using an unsupervised approach to recreate the input vector rather than to assign a class label to it. Usually, the number of hidden units is smaller than the input/output layers, which achieve encoding of the data in a lower dimensional space and extract the most discriminative features.
HIDDEN LAYER
INPUT LAYER
OUTPUT LAYER
CNNs have been proposed to process efficiently imagery data. The name of these networks comes from the convolution operator that provides an easy way to perform complex operations using convolution filter. CNNs use locally connected neurons that represent data specific kernels. The main advantage of a CNN is that during back-propagation, the network has to adjust a number of parameters equal to a single instance of the kernel which drastically reduces the connections from the typical neural network. The concept of CNN is inspired by the neurobiological model of the visual cortex and can be briefly summarized as a sequence of convolution and subsampling of the image until high level features can be extracted.
Recurrent neural nets (RNN)
RNN is a neural network that contains hidden units capable of analysing streams of data. Since RNN suffers from the vanishing gradient and exploding gradient problems, a variation called Long Short-Term Memory units (LSTMs) was proposed in 1997 to solve this problem. Specifically, LSTM is particularly suitable for applications where there are very long time lags of unknown sizes between important events.
RNN and LSTM share the same weights across all steps that greatly reduce the total number of parameters that the network needs to learn. RNNs have shown great successes in many Natural Language Processing tasks such as language modelling, bioinformatics, speech recognition and generating image description.
Ot-2
Ot-1
Ot
Ot+1
OUTPUT STREAM
St-2
St-1
St
St+1
MEMORY
Yt-2
Yt-1
Yt1
Yt+1
INPUT STREAM
1. HARDWARE FOR AI
In 1965, Gordon Moore observed that the number of transistors, in a dense integrated circuit, doubles
approximately every year. Ten years later, he revised his forecast, updating his prediction to the number doubling every two years. Moore's prediction has been accurate for several decades and has been used in the semiconductor industry to guide long-term planning. In 2015, Moore realised that the rate of progress in the hardware would reach saturation and the transistors would arrive at the limits of miniaturisation at the atomic level. Experts estimate that Moore’s law could end in 2025. Today, his prediction is still valid and the number of transistors is increasing even if, after 2005, the frequency and the power started to reduce, leading to a core scaling rather than a frequency improvement. Therefore, since 2005, we are no longer getting faster computers, but the hardware is designed in a multi-core manner. To take full advantage of this different hardware implementation, the software has to be written in a multi-threaded manner too. In future, experts believe that revolutionary technologies may help sustain Moore's law. One of the key challenges will be the design of gates in nanoscale transistors and the ability of controlling the current flow as, when the device dimension shrinks, the connection between transistors becomes more difficult.
1. ROBOTICS AND AI
Building on the advances made in mechatronics, electrical engineering and computing, robotics is developing increasingly sophisticated sensorimotor functions that give machines the ability to adapt to their ever-changing environment. Until now, the system of industrial production was organized around the machine; it is calibrated according to its environment and tolerated minimal variations.
Today, it can be integrated more easily into an existing environment. The autonomy of a robot in an environment can be subdivided into perceiving, planning and execution (manipulating, navigating, collaborating). The main idea of converging AI and Robotics is to try to optimise its level of autonomy through learning. This level of intelligence can be measured as the capacity of predicting the future, either in planning a task, or in interacting (either by manipulating or navigating) with the world. Robots with intelligence have been attempted many times. Although creating a system exhibiting human-like intelligence remains elusive, robots that can perform specialized autonomous tasks, such as driving a vehicle [52], flying in natural and man-made environments [53], swimming [54], carrying boxes and material in different terrains [55], pick up objects [56] and put them down [57]
do exist today.
Another important application of AI in robotics is for the task of perception. Robots can sense the environment by means of integrated sensors or computer vision. In the last decade, computer systems have improved the quality of both sensing and vision. Perception is not only important for planning but also for creating an artificial sense of self-awareness in the robot. This permits supporting interactions with the robot with other entities in the same environment. This discipline is known as social robotics. It covers two broad domains: human-robot interactions (HCI) and cognitive robotics.
The vision of HCI it to improve the robotic perception of humans such as in understanding activities [58], emotions [59], non-verbal communications [60] and in being able to navigate an environment along with humans [61]. The field of cognitive robotics focuses on providing robots with the autonomous capacity of learning and acquiring knowledge from sophisticated levels of perception based on imitation and experience. It aims at mimicking the human cognitive system, which regulates the process of acquiring knowledge and understanding, through experience and sensorisation [62]. In cognitive robotics, there are also models that incorporate motivation and curiosity to improve the quality and speed of knowledge acquisition through learning
[63, 64].
AI has continued beating all records and overcoming many challenges that were unthinkable less than a decade ago. The combination of these advances will continue to reshape our understanding about robotic intelligence in many new domains. Figure 9 provides a timeline of the milestone in robotics and AI.
1. PROGRAMMING LANGUAGES FOR AI
Programming languages played a major role in the evolution of AI since the late 1950s and several teams carried out important research projects in AI; e.g. automatic demonstration programs and game programs (Chess, Ladies) [65]. During these periods researchers found that one of the special requirements for AI is the ability to easily manipulate symbols and lists of symbols rather than processing numbers or strings of characters. Since the languages of the time did not offer such facilities, a researcher from MIT, John MacCarthy, developed, during 1956-58, the definition of an ad-hoc language for logic programming, called LISP (LISt Processing language).
Since then, several hundred derivative languages,
so-called "Lisp dialects", have emerged (Scheme, Common Lisp, Clojure); Indeed, writing a LISP interpreter is not a hard task for a Lisp programmer (it involves only a few thousand instructions) compared to the development of a compiler for a classical language (which requires several tens of thousands of instructions). Because of its expressiveness and flexibility, LISP was very successful in the artificial intelligence community until the 1990s.
Another important event at the beginning of AI was the creation of a language with the main purpose of expressing logic rules and axioms. Around 1972 a new language was created by Alain Colmerauer and Philippe Roussel named
Prolog (PROgramming in Logic). Their goal was to create a programming language where the expected logical rules of a solution can be defined and the compiler automatically transforms it into a sequence of instructions. Prolog is used in AI and in natural language processing. Its rules of syntax and its semantics are simple and considered accessible to non-programmers. One of the objectives was to provide a tool for linguistics that was compatible with computer science.
In the 1990s, the machine languages with C / C ++ and Fortran gained popularity and eclipsed the use of LISP and Prolog. Greater emphasis was placed on creating functions and libraries for scientific computation on these platforms and were used for intensive data analysis tasks or artificial intelligence with early robots. In the middle of the 1990s, the company Sun Microsystems, started a project to create a language that solved secutiry flaws, distributed programming and multi-threading of C++. In addition, they wanted a platform that could be ported to any type of device or platform. In 1995, they presented Java, which took the concept of object orientation much further than C++.
Equally, one of the most important additions to Java was the Java VM (JVM) which enabled the capability of running the same code in any device regardless of their internal technology and without the need of pre-compiling for every platform. This added new advantages to the field of AI that were be introduced in devices such as cloud servers and embedded computers. Another important feature of Java was that it also offered one of the first frameworks, with specific tools for the internet, bringing the possibility of running applications in the form of java applets and javascripts (i.e. self-executing programs) without the need of installation. This had an enormous impact in the field of AI and a set the foundation in the fields of web 2.0/3.0 and the internet of things (IoT).
However, the development of AI using purely procedural languages was costly, time-consuming and error prone. Consequently, this turned the attention into other multi- paradigm languages that could combine features from functional and procedural object-oriented languages.
Python, although first published in 1991, started to gain popularity as an alternative to C/C++ with Python 2.2 by 2001. The Python concept was to have a language that could be as powerful as C/C++ but also expressive and pragmatic for executing "scripts" like Shell Script. It was in 2008, with the publication of Python 3.0, which solved several initial flaws, when the language started to be considered a serious contender for C++, java and other scripting languages such as Perl.
Since 2008, the Python community has been trying to catch up with specific languages for scientific computing, such as Matlab and R. Due to its versatility, Python is now used frequently for research in AI. However, although python has some of the advantages of functional programming, run-time speeds are still far behind other functional languages, such as Lisp or Haskell, and even more so from C/C++.
In addition, it lacks of efficiency when managing large amounts of memory and highly-concurrent systems.
From 2010 and mostly driven by the necessity of translating AI into commercial products, (that could be used by thousands and millions of users in real time), IT corporations looked for alternatives by creating hybrid languages, that combined the best from all paradigms without compromising speed, capacity and concurrency. In recent years, new languages such as Scala and Go, as well as Erlang or Clojure, have been used for applications with very high concurrency and parallelization, mostly on the server side.
Well-known examples are Facebook with Erlang or Google with Go. New languages for scientific computation have also emerged such as Julia and Lua.
Although functional programming has been popular in academia, its use in industrial settings has been marginal and mainly during the times when “expert systems” were at their peak, predominantly during the 1980s. After the fall of expert systems, functional programing has, for many years, been considered a failing relic from that period. However, as multiprocessors and parallel computing are becoming more available, functional programming is proving to be a choice of many programmers to maximise functionality from their multicore processors. These highly expensive computations are usually needed for heavy mathematical operations or pattern matching, which constitute a fundamental part of running an AI system. In the future, we will see new languages that bring simplifications on existing functional languages such as Haskell and Erlang and make this programming paradigm more accessible. In addition, the advent of the internet-of-things (IoT) has drawn the attention to the programming of embedded systems. Thus, efficiency, safety and performance are again matters for discussion.
New languages that can replace C/C++ incorporating tips from functional programming (e.g. Elixir) will become
increasingly popular. Also, new languages that incorporate simplifications as well as a set of functions from modern imperative programming, while maintaining a performance like C/C++ (e.g. Rust), will be another future development.
1. IMPACT OF MACHINE VISION
Machine vision integrates image capture systems with computer vision algorithms to provide automatic inspection and robot guidance. Although it is inspired by the human vision system, based on the extraction of conceptual information from two-dimensional images, machine vision systems are not restricted to 2D visible light. Optical sensors include single beam lasers to 3D high definition Light Detection And Ranging (LiDAR) systems, also known as laser scanning 2D or 3D sonar sensors and one or multiple 2D camera systems. Nevertheless, most machine vision applications are based on 2D image-based capture systems and computer vision algorithms that mimic aspects of human visual perception. Humans perceive the surrounding world in 3D and their ability to navigate and accomplish certain tasks depends on reconstructing 3D information from 2D images that allows them to locate themselves in relation to the surrounding objects. Subsequently, this information is combined with prior knowledge in order to detect and identify objects around them and understand how they interact. Scene reconstruction along with object detection and recognition are the main sub-domains of computer vision.
Regardless of the imaging sensors used, the most common approaches of reconstructing 3D information are normally based on either time-of-flight techniques, multi-view geometry and/or on photometric stereo. The former is used in laser scanners to estimate the distance between the light source and the object based on the time required for the light to reach the object and return back. Time-of-flight approaches are used to measure distances in kilometres and they are accurate to a millimetre scale, since they are limited by the ability to measure time. On the other hand, multi-view geometry problems include ‘structure’ problems, ‘stereo correspondence’ problems and ‘motion’ problems. Recovery of the 3D ‘structure’ implies that given 2D projections of the same 3D point, in two or more images, the 3D coordinates of the point are estimated based on triangulation. ‘Stereo correspondence’ refers to the problem of finding the image point that corresponds to a point from another 2D view.
Finally, ‘motion’ refers to the problem of recovering the camera coordinates given a set of corresponding points in two or more image views. 3D laser scanners based on triangulation can reach micrometre accuracy but their range is constrained to a few meters. Several sub-problems such as ‘structure from motion’ uses multi-view geometry principles to extract corresponding points between 2D views of the same object and reconstruct its shape.
Stereo-vision assumes the robust extraction of corresponding salient points/features across images, the so-called interest point detection. These features should be invariant to photometric transformation such as changes in the lighting conditions and covariant to geometric transformations.
For over two decades researchers have proposed several approaches. The Scale-invariant feature transform (SIFT) extracts features that are invariant to scale, rotation and translation transformations and robust to illumination variations and moderate perspective transformations.
Since its introduction in 1999-2004, it has been successful in several vision applications, including object recognition, robot localisation and mapping.
Representing and recognising object categories have proven much harder problems to generalise and solve, compared with 3D reconstruction, since there are thousands of objects that can belong to an arbitrary number of categories simultaneously. Several ideas about object detection are related to Gestalt psychology, which is a theory of mind
with relation to visual perception. A major aspect of the theory is about grouping entities together based on their proximity, similarity, symmetry, common fate, continuity and so on. From the 1960s to early 1990s, research in object recognition was centred on geometric shapes. This was a bottom-up process, which uses a small number of primitive 3D dimensional objects that are assembled together in various configurations to form complex objects. In the 1990s, appearance-based models were explored, which were based on manifold learning of the object appearance parameterised by the pose and illumination [66]. These techniques are not robust to occlusion, clutter and deformation. By the mid-late 1990s, sliding window approaches were designed that classify whether an object is found for each instance of a sliding window across an image [67]. The main challenges were how to design features that represent appropriately the appearance of the object and how to efficiently search a large number of positions and scales. Local features approaches were also developed and they aimed towards those which were invariant to image scaling, geometric transformations and illumination changes [68]. In the early 2000s 'parts-and- shape' models along with 'bags of features' were suggested. Parts-and-shape models represent complex objects using combinations of multi-scaled deformable objects [69].
On the other hand, bags of features methods, represent visual features as words and relate object recognition and image classification to the expressive power of natural language processing approaches [70].
Machine learning in object recognition facilitated a shift, from solving a problem based on mathematical modelling alone, to learning algorithms based on real-data and statistical modelling. A major breakthrough in object recognition and classification came in 2012 with the emergence of deep neural networks and the availability of large labelled image databases, such as ImageNet. Compared to classical object recognition methods, which depend on feature extraction followed by feature matching methodologies, deep learning has the advantage of encoding both feature extraction and image classification via the structure of a neural network.
The superb performance of deep neural networks resulted in an increase of image classification from 72% in 2010 to 96% in 2015, which outperforms human accuracy and has had a significant impact in real-life applications [71]. Both Google and Baidu updated their image search capabilities based on the Hinton’s deep neural network architecture. Face detection has been introduced in several mobile devices and Apple even created an app to recognise pets. The accuracy of these models in object recognition and image classification exceed human-level accuracy and spread waves of technological changes across the industry.
Post a Comment