The concept of Artificial Intelligence (AI) has changed over time, but the core theme has always been of machines capable of thinking like humans. It makes more sense to use human intelligence as a goal since human beings have proven themselves to be uniquely capable of interpreting information from the environment through abstract reasoning, mental representation, problem solving, and decision making. Moreover, humans have the ability to learn and adapt to meet the demands of changing environments effectively. AI can then be conceptualized as simulating these capabilities by using the digital processes of computers. In a groundbreaking paper, “Computing Machinery and Intelligence” published in 1950 in ‘Mind’, Alan Turing considered the mathematical possibility of artificial intelligence, with the question that, given a computer with adequate storage, suitably increasing its speed of action, and providing it with an appropriate program (the three basic requirements of AI), “Can machines do what we (as thinking entities) can do?” He setup an intelligence comparison test for that purpose, which is now famously referred to as the “Turing Test”: if human judges cannot effectively distinguish between a computer and a human through their text responses in conversations, then it must be concluded that the computer is intelligent. In the summer of 1956, John McCarthy, a professor of mathematics at Dartmouth College, organized a conference called ‘Summer Research Project on Artificial Intelligence’ to provide a forum for researchers to discuss ways in which computers could be programmed to carry out intelligent behavior. He coined the term ‘artificial intelligence’ for that machine, defining Artificial Intelligence as, “making a machine behave in ways that would be called intelligent if a human were so behaving”. During the conference, Herb Simon and Alan Newell from Carnegie Institute of Technology succeeded in creating and demonstrating the artificial intelligence machine program that McCarthy had envisioned. This primitive program, which they called ‘logic theorist’, used human-like reasoning processes to solve problems. Interestingly, none of the scientists attending at that time had any idea that the conferences would mark the birth of cognitive science, a new way of thinking about the human mind. In a 2004 paper, John McCarthy defined artificial intelligence as, “the science and engineering of making intelligent machines, especially intelligent computer programs.” Stuart Russell and Peter Norvig in, ‘Artificial Intelligence: A Modern Approach’ (first published in 1995); a leading textbook in the study of AI, delved into the four potential goals or definitions of AI that differentiate computer systems on the basis of thinking or acting humanly versus rationally:
- Human approach:
- Systems that think like humans
- Systems that act like humans
- Ideal approach:
- Systems that think rationally
- Systems that act rationally
Even as AI enthusiasts started believing that soon computers would be able to carry on conversations, translate languages, interpret pictures, and reason like people, computer scientists found it incredibly difficult to create intelligent machines despite well-funded global efforts over several decades across the late twentieth century. One important problem that the early day AI researchers had was the lack of adequate computing speed, storage, and programming capabilities (the three basic requirements of AI) to facilitate AI. However, exponential gains in computer processing power and storage ability, in the late 1990s and early 21st century, allowed companies to store and crunch large quantities of data. The availability of vast amounts of data with digitalization provided the necessary push for AI. The data available from, for example, social media and machine data generated by connected industrial machinery enabled computers to learn more efficiently and make better decisions.
Intelligence is often referred to as the ability of an entity to solve problems and achieve goals in diverse environments. In the context of AI, S. Legg and M. Hutter (2007) summarized 70 definitions from the literature into a single statement: “Intelligence measures an agent’s ability to achieve goals in a wide range of environments.” In fact, research and development work in AI today is split between two branches: one is termed as ‘applied AI’, ‘weak AI’, or ‘narrow AI’, which uses the principles of simulating human thought processes to carry out specific tasks within a particular context; the other is known as ‘Artificial general intelligence (AGI)’, or ‘general AI’ or ‘strong AI’ or ‘human-level AI’, which seeks to develop machine intelligences that can flexibly perform various tasks across many different environments, more like a human can. Artificial general intelligence (AGI) is a theoretical form of AI where a machine would have intelligence equal to humans, be self-aware (conscious), endowed with the ability to solve problems, learn, and adapt to changing environments. Another level of AI is referred to as ‘Artificial Super Intelligence (ASI)’, which would surpass the intelligence and abilities of the human brain in many aspects. AI has basically achieved far greater commercial success in specific ‘narrow’ problems, where specialized applied AI systems are now being used extensively in practice. Research in these areas has easily found funds across the spectrum of uses. This is already providing breakthroughs in fields of study such as in speech recognition and voice search. In medicine and healthcare, it is being used for radiology imaging and diagnosis of patients based on genomic data. In the financial world, it finds uses in applications spanning across from fraud detection, to providing personalized advice, and automated stock trading for optimizing stock portfolios. In industry, it is employed for uses ranging from improving customer service with chatbots to cross-selling products with recommendation engines. In manufacturing, it is used in automated robots, as well as for enabling predictive maintenance by identifying vibrations or unusual sounds that ensures maintenance is performed on hardware before it malfunctions. While in the consumer world, this technology is being embedded into everyday human lives with virtual assistants like Apple’s Siri and Amazon’s Alexa, and other Internet of Things (IoT) home devices. However, even advanced systems such as virtual assistants and robots are far ‘narrower’ in breadth of performance as compared to the human brain. Today’s AI technologies are all categorized as narrow AI. They’re ‘narrow’ in the sense that today’s AI applications always specialize in solving specific problems in defined environments based on rules established specifically for a particular task. In many areas, narrow AI has already surpassed specific capabilities of human beings. So, although a chess program may be able to continually optimize its game strategy to defeat any human, it would not be capable of analyzing genomic data unless specifically designed to do so. Strong (broader) AI is differentiated by the ability to transfer knowledge and skills from one context to another and to adapt flexibly to a variety of contexts, even new and unfamiliar ones. It is also able to make decisions on its own, and to solve many different problems; as also, to communicate with other machines and human beings. For example, virtual assistants like Siri or Alexa with strong AI wouldn’t be just reactive responders but proactive advisors that would be able to not only predict our needs but also respond to them.
Arthur Lee Samuel, an American pioneer in the field of computer gaming and artificial intelligence, coined the term ‘Machine Learning’ in the year 1959. He defined Machine Learning as a “Field of study that gives computers the capability to learn without being explicitly programmed”. Machine learning represents the current methodology for achieving AI. With programming, the input data and a tested program is fed into a machine to generate the output. Whereas, in machine learning, the input data along with the desired output or the goal is fed into the machine, which then learns from the input datasets and works out the prediction program or algorithm to achieve the goal all by itself. This algorithm is then used to predict the outputs from new input sets of data. The accuracy of these models would depend upon the quality and amount of input data. Large quantities of quality data will help build a better model for predicting the output more accurately. In a supervised machine learning approach, the learning process requires a training dataset, which includes both a labeled dataset of desired outputs and an input dataset, which the machine uses to learn by mapping or comparing these in order to build an algorithm of the machine learning solution. By continuously comparing the two input datasets and correcting the model, an algorithm with the mapping function is developed that has ‘learnt’ to classify the input dataset similar to the training dataset. This ‘trained’ model can then be used for predictions, to classify or map any new input datasets. For example, a supervised machine learning model would be fed labeled training datasets of pictures of two breeds of dogs. The machine will compare pictures of dogs in an input dataset to determine a feature algorithm for classifying this dataset into the two breeds, similar to the training dataset. It is then said to have ‘learned’ to classify the pictures of the two breeds of dogs in the form of an algorithm. This model algorithm can then be used for predictions, to classify any input dataset of unclassified dog pictures into an output dataset of pictures with the two breeds of dogs classified. This model provides ‘weak’ or ‘narrow’ AI solutions, applicable only for a specific well-defined problem. On the other hand, if the machine is only fed input datasets to find patterns in the data on its own, without being provided any references for the kinds of patterns to look for, the learning process is referred to as unsupervised machine learning. This model would only use unclassified dog pictures as inputs, for the example given earlier. The machine will then learn to classify the dog pictures on the basis of the differences between the dog’s features in the pictures on its own, to create a classification algorithm for predictive use on new datasets. Here the machine is provided input datasets with a goal to find patterns, without the related benchmarking labeled training datasets for comparison. Unsupervised learning approach yields ‘stronger’ AI since the machine is working on its own to solve a not very well-defined problem, but still provides a ‘narrow’ solution to a specific problem.
Machine learning models are limited to shallow learning that reaches a plateau after a certain level of data, with new data additions providing only marginal improvements to its learning. To overcome this, the application of neuroscience has led to the development of artificial neural networks. Neural networks are mathematical machine models that try to replicate the human brain’s neural circuits to learn by screening and gathering information from data automatically. The structure and functioning of neural networks are very loosely based on the network connections between neurons in the human brain. An Artificial Neural Network (ANN) is made up of interconnected layers of ‘neural’ algorithms which feed data into each other. The process of training these neural networks to carry out specific tasks involves attributing importance to data as it passes between layers. The data continues to be modified as it moves across the layers until the output from the neural network is very close to what is desired. At this point the network would have ‘learned’ how to carry out a particular task. Deep Learning is a machine learning technique that uses neural networking to learn. Deep learning can be applied to different learning archetypes, including supervised learning, reinforcement learning, as well as unsupervised learning. In deep learning, neural networks apply deep networks with large number of sizeable layers that are trained using massive amounts of data. At its core, deep learning relies on iterative methods to teach machines to imitate human intelligence. An artificial neural network carries out this iterative method through several hierarchical levels. The initial levels help the machines learn simple information, and as the levels increase, the information keeps building up. With each new level, the machine picks up further information and combines it with what it had learnt in the last level. This information passes through several hierarchies and can be compared to human logical thinking. Pattern recognition is an essential part of deep learning. Through deep learning, machines can use images, text or audio files to perform any task in a human-like manner. The desired output could be, for example, visual recognition of faces or detected credit card frauds. Since deep learning involves end-to-end learning, the more data it uses, the more precise and accurate the results are. Although supervised machine learning requires large datasets of labelled data to be accurate, a semi-supervised learning approach with Generative Adversarial Networks (GANs) can be used to train the machine for tasks with small amounts of labelled data alongside large amounts of unlabeled data. Generative modeling is an unsupervised learning task in machine learning that involves framing the problem as a supervised learning problem with two sub-models automatically discovering and learning the similarities or patterns in input data in such a way that the model can be used to generate or output new synthetic examples that appear to have been drawn from the original dataset. GANs are widely used in image, video, and voice generation.
Reinforcement Learning (RL) theory (proposed originally by B. F. Skinner in 1938, an American psychologist, as ‘Operant Conditioning’) states that, if the external environment is designed so as to motivate an organism to perform desired actions with rewards, the organism’s behavior tends to be repeated, leading to learning. On the other hand, behavior of the organism, which is punished, will occur less frequently. Reinforcement learning of machine learning models or agents, similarly, involves learning, in a game-like situation within an uncertain and complex environment, to achieve a goal in the form of a solution to a problem posed to it. The two main components of RL are the environment, which contains the problem to be solved, and, rewards or penalties for actions of the model or agent, which represents the learning algorithm. The model itself figures out how to perform the task by performing a sequence of desired actions in an environment, through decisions employing trial and error when faced with choices, to maximize the rewards until it arrives at the best possible outcome. Its goal is to maximize the total reward without being informed how to do so. For AI applications, a reinforcement learning algorithm can learn by using thousands of parallel gameplays given sufficiently powerful computer infrastructure. RL is also used to train autonomous robots in real-world environments. Google DeepMind’s Deep Q-network has been trained to perform better than humans in a variety of classic video games with RL. Self-driving car companies use algorithms based on reinforcement learning to develop the control systems by training machines to take decisions through rewarding and punishing interactions within the environment. Evolutionary Computation (EC) is another research area of AI, which is based on Darwin’s theory of natural selection. In this, algorithms undergo genetic mutations and combinations between generations in an attempt to evolve the optimal solution to a given problem. An EC algorithm starts by creating a population consisting of algorithms that represent solutions to the problem. These algorithms are then evaluated with a fitness function, having operators inspired by natural evolution, such as crossover, mutation, selection, and reproduction. The fitness function evaluates how well the algorithms solve, or come close to solving, the problem. This generates a new population of evolved algorithms, with some algorithms getting eliminated. This process continues until a termination criterion, for example, of reaching the number of generations defined, is reached to stop the algorithm. The best algorithm with the highest fitness value is selected as the solution. This approach is even being used to help design AI models to, effectively, use AI itself to help build better AI. This use of evolutionary algorithms to optimize neural networks is called ‘neuroevolution’, and is most commonly applied in game playing and evolutionary robotics The technique was showcased by Uber AI Labs, which released research on using genetic algorithms to train deep neural networks for reinforcement learning problems. Then, there are also ‘expert’ systems, where computers are designed to solve complex problems by reasoning through bodies of knowledge, allowing the machine to imitate the behavior of a human expert in a specific domain. An example use of these knowledge-based systems is real time process control. Eliza Kosoy, a researcher in MIT’s Center for Brains, Minds, and Machines is researching to better understand certain human learning capabilities that machines are as yet lacking: Intuitive physics refers to the way in which humans are able to predict certain dynamic changes in their physical environment, and then respond accordingly to these changes. For example, the ability to gauge the trajectory of a falling object, in order to move away, to avoid getting hit. Also, one-shot learning is the ability to learn object categories from only a few examples. For instance, children can distinguish between different fruits after only a few observations. In comparison, the best algorithms today would need exposure to large datasets in order to distinguish between the fruits. Kosoy is studying the quick learning ability of children in order to build faster machine learning models that are able to process with much less data.
Leading us into AGI are the tech giants, who are fiercely competing for the next breakthrough. Originally designed to defeat two grandmasters on the TV quiz show Jeopardy!, IBM’s Watson cognitive computing platform uses high-level simulations of human neurological processes, by powering advances in computer-based sensing, understanding, and action, to carry out an ever-growing range of tasks without being specifically taught how to do them. Google’s DeepMind is looking to combine machine learning and the pursuit of neuroscience to develop more general and capable problem-solving systems for AGI. Google’s DeepMind, in a paper to the peer-reviewed Artificial Intelligence journal titled “Reward is Enough” proposed that reinforcement learning will one day allow for instantaneous calculation and perfect memory, leading to an artificial intelligence that would outperform humans at nearly every cognitive task. Google Brain has announced a deep-learning computer vision model containing two billion parameters. Microsoft has invested in OpenAI, an independent AI research laboratory, to develop a hardware and software platform within Microsoft Azure which they expect to scale to AGI. OpenAI has released a series of powerful natural language generation models under the name GPT (Generative Pre-trained Transformer). In 2020, they released GPT-3, which demonstrated that after an initial prompt, it can recognize and replicate patterns of words to work out what comes next. But although GPT-3 can predict what the next word in a sentence should be with uncanny accuracy, it has no sense of meaning. Researchers from the Beijing Academy of Artificial Intelligence (BAAI) in China recently introduced Wu Dao 2.0, a multimodal-AI system with 1.75 trillion parameters. Like GPT-3, Wu Dao uses multimodal and multitasking models towards AGI. It can perform natural language processing, text generation, image recognition, and image generation tasks.
Human beings are able to use their intelligence in a variety of contexts. The human brain is extremely flexible and can adapt intuitively to unpredictable environments. Certain characteristics of the human brain are very difficult to simulate: humans are creative, curious and endowed with social skills, all of which continues to set humans apart from even the most intelligent computer. Even though computers have created paintings and literature, these don’t compare with the innumerable examples of works created by boundless human creativity. Doubts abound about whether AI could ever be curious enough, just like humans, to wonder about nature, the universe or even themselves. For achieving AGI, machine learning requires sufficient number of self-optimizing powerful and dynamic algorithms with the ability to recognize relevant patterns in the sensed data, and to learn and adapt continually to changing conditions. AGI also needs knowledge and experience in the form of large quantities of data in order to select and apply the optimal algorithms for solving innumerable different problems it would face. Another aspect of human intelligence is emotion: Can Siri, Alexa or Assistant be expected to emote with humans during celebrations or grief in the household? Hence, Generalized AI is not anywhere in the horizon just yet. A complete simulation of the human brain would require more understanding of the human brain, as also, much more computing power than is presently available. Geoffrey Hinton, University of Toronto professor, who is a pioneer of deep learning and a Turing Award winner, notes: “There are one trillion synapses in a cubic centimeter of the brain. If there is such a thing as general AI, [the system] would probably require one trillion synapses.” Conventional computer systems are already reaching the limits of their capabilities for strong AI.
The first generation of AI was based on rules and used classical logic to obtain solutions for specific, narrowly-defined problem domains like, for example, monitoring processes and improving efficiency. The current generation of AI is involved in sensing and perception, by using deep-learning networks, for image recognition and machine vision, for example. The future generation would extend AI into areas of reasoning on the basis of knowledge and memory, and the use of machine cognition for interpretation and autonomous adaptation. Research institutions and companies all over the world are working on entirely new computing technologies that could bring AI to the next level. The biggest breakthroughs in machine learning and deep learning have been driven not only by the easy availability of data, but even more so, by an explosion in parallel computing power using clusters of graphics processing units (GPUs) to train machine-learning systems. Not only do these clusters offer more powerful systems for training machine-learning models, but these are now widely available as cloud services over the internet. All of the major cloud platforms like Amazon Web Services, Microsoft Azure and Google Cloud Platform provide access to GPU arrays for training and running machine-learning models. Demand for increased chip performance and energy efficiency continues to rise, especially in the era of hybrid cloud, AI, and the Internet of Things. Google lets users use its Tensor Processing Units {TPUs}, which are custom chips whose design is optimized for training and running machine-learning models. On May 6, 2021, IBM unveiled the development of the world’s first chip with 2 nanometer (nm) nanosheet technology with 50 billion transistors on a fingernail-sized chip. This makes chips smaller, faster, more reliable, and more efficient, while providing more options to improve capabilities for leading edge workloads like AI and cloud computing. A new generation of computer chip technology known as neuromorphic processors are being designed to more efficiently run brain-simulator code. Neuromorphic chips use less energy and deliver better performance than conventional CPU chips because they are designed differently. Instead of passing information from one transistor to the next in a series of billions of transistors, neuromorphic chips contain a million “neurons” that can pass information in any direction amongst the neurons via 256 million connections, called “synapses.” Traditional CPUs process instructions based on ‘clocked time’, wherein information is transmitted at regularly timed intervals. In comparison, neuromorphic chips, packed with digital equivalents of neurons, communicate, without the constriction of clocked time, in parallel, using “spikes” or bursts of electric current that can be sent whenever needed. Just like human brains, a neuromorphic chip’s neurons communicate by determining from the incoming spike whether to send current out to the next neuron. A research paper by Intel scientist Charles Augustine predicts that neuromorphic chips will soon be able to handle artificial intelligence tasks such as cognitive computing, adaptive artificial intelligence, sensing data, and associate memory. Developments in quantum computing look promising for AI, in processing huge amounts of complex datasets, and developing algorithms to allow for better learning, reasoning and understanding. Dr. Jay Gambetta, who is an IBM Fellow and vice president of IBM Quantum, says, “Quantum computing is a new kind of computing, using the same physical rules that atoms follow in order to manipulate information. At this fundamental level, quantum computers execute quantum circuits—like a computer’s logical circuits, but now using the physical phenomena of superposition, entanglement, and interference to implement mathematical calculations out of the reach of even our most advanced supercomputers.”
Although Artificial Intelligence technologies have made substantial progress and created impacts in industry and daily life, concerns about privacy, bias, inequality, trust, safety, interpretability and accountability of AI have arisen. With AI systems being able to create near-real fake images, videos, conversations, and other content, it’s becoming difficult to distinguish real from the machine-generated. As AI systems get smarter, they become more efficient at identifying and targeting vulnerabilities in systems and can make threats to infrastructure harder to detect. Deep learning outputs are obtained through a ‘black box’ process which cannot be understood and explained. This creates issues when we rely on this technology to make critical decisions such as diagnostics, loan approvals, promotions and recruitments, which impact people’s lives. Explainable AI is necessary to develop trust in AI systems. Many influential thinkers have recognized the need for sound explainability to be integrated with deep learning. The “Global AI Adoption Index 2021,” conducted by Morning Consult on behalf of IBM, revealed that “91 percent of businesses using AI say their ability to explain how it arrived at a decision is critical. While global businesses are now acutely aware of the importance of having trustworthy AI, more than half of companies cite significant barriers in getting there including lack of skills, inflexible governance tools, biased data and more.” According to a Harvard Business Review article, “Biases in Your Organization” by Jennifer M. Logg, “….algorithms are tools. People build them, determine if their output is accurate, and decide when and how to act on that output. Data can provide insights, but people are responsible for the decisions made based on them.” It will be critically important to deal responsibly with new technologies like AI by establishing strict ethical standards and creating policies accepted throughout the world for managing these technologies. This understandable concern has led to the coming together of a number of tech giants including Google, IBM, DeepMind, Microsoft, Facebook and Amazon, in The Partnership on AI (publicly announced on September 28, 2016), which has been “established to study and formulate best practices on AI technologies, to advance the public’s understanding of AI, and to serve as an open platform for discussion and engagement about AI and its influences on people and society.” And to, “address such areas as fairness and inclusivity, explanation and transparency, security and privacy, values and ethics, collaboration between people and AI systems, interoperability of systems, and of the trustworthiness, reliability, containment, safety, and robustness of the technology.” Similarly, launched in June 2020, The Global Partnership on Artificial Intelligence (GPAI) is a multi-stakeholder initiative that “provides a mechanism for sharing multidisciplinary research and identifying key issues among AI practitioners, facilitating international collaboration and promoting the adoption of trustworthy AI.”
There’s no doubt that AI systems will continue to make life more comfortable, enjoyable and exciting for humans. But would a machine ever be our emotional companion who ‘enjoys’ with us or expresses grief like a human being? It’s clear that the human brain is a wondrous thing that is capable of creating machines that mirror us, and even better us, in more and more ways. The question, however, remains whether humans would be able to create machines that are not only smarter than us and humane like us, but also have beautiful minds.