Future Goals in the AI Race: Explainable AI and Transfer Learning

Recent years have seen breakthroughs in neural network technology: computers can now beat any living person at the most complex game invented by humankind, as well as imitate human voices and faces (both real and non-existent) in a deceptively realistic manner. Is this a victory for artificial intelligence over human intelligence? And if not, what else do researchers and developers need to achieve to make the winners in the AI race the “kings of the world?”

Background

Over the last 60 years, artificial intelligence (AI) has been the subject of much discussion among researchers representing different approaches and schools of thought. One of the crucial reasons for this is that there is no unified definition of what constitutes AI, with differences persisting even now. This means that any objective assessment of the current state and prospects of AI, and its crucial areas of research, in particular, will be intricately linked with the subjective philosophical views of researchers and the practical experience of developers.

In recent years, the term “general intelligence,” meaning the ability to solve cognitive problems in general terms, adapting to the environment through learning, minimizing risks and optimizing the losses in achieving goals, has gained currency among researchers and developers. This led to the concept of artificial general intelligence (AGI), potentially vested not in a human, but a cybernetic system of sufficient computational power. Many refer to this kind of intelligence as “strong AI,” as opposed to “weak AI,” which has become a mundane topic in recent years.

As applied AI technology has developed over the last 60 years, we can see how many practical applications – knowledge bases, expert systems, image recognition systems, prediction systems, tracking and control systems for various technological processes – are no longer viewed as examples of AI and have become part of “ordinary technology.” The bar for what constitutes AI rises accordingly, and today it is the hypothetical “general intelligence,” human-level intelligence or “strong AI,” that is assumed to be the “real thing” in most discussions. Technologies that are already being used are broken down into knowledge engineering, data science or specific areas of “narrow AI” that combine elements of different AI approaches with specialized humanities or mathematical disciplines, such as stock market or weather forecasting, speech and text recognition and language processing.

Different schools of research, each working within their own paradigms, also have differing interpretations of the spheres of application, goals, definitions and prospects of AI, and are often dismissive of alternative approaches. However, there has been a kind of synergistic convergence of various approaches in recent years, and researchers and developers are increasingly turning to hybrid models and methodologies, coming up with different combinations.

Since the dawn of AI, two approaches to AI have been the most popular. The first, “symbolic” approach, assumes that the roots of AI lie in philosophy, logic and mathematics and operate according to logical rules, sign and symbolic systems, interpreted in terms of the conscious human cognitive process. The second approach (biological in nature), referred to as connectionist, neural-network, neuromorphic, associative or subsymbolic, is based on reproducing the physical structures and processes of the human brain identified through neurophysiological research. The two approaches have evolved over 60 years, steadily becoming closer to each other. For instance, logical inference systems based on Boolean algebra have transformed into fuzzy logic or probabilistic programming, reproducing network architectures akin to neural networks that evolved within the neuromorphic approach. On the other hand, methods based on “artificial neural networks” are very far from reproducing the functions of actual biological neural networks and rely more on mathematical methods from linear algebra and tensor calculus.

Are There “Holes” in Neural Networks?

In the last decade, it was the connectionist, or subsymbolic, approach that brought about explosive progress in applying machine learning methods to a wide range of tasks. Examples include both traditional statistical methodologies, like logistical regression, and more recent achievements in artificial neural network modelling, like deep learning and reinforcement learning. The most significant breakthrough of the last decade was brought about not so much by new ideas as by the accumulation of a critical mass of tagged datasets, the low cost of storing massive volumes of training samples and, most importantly, the sharp decline of computational costs, including the possibility of using specialized, relatively cheap hardware for neural network modelling. The breakthrough was brought about by a combination of these factors that made it possible to train and configure neural network algorithms to make a quantitative leap, as well as to provide a cost-effective solution to a broad range of applied problems relating to recognition, classification and prediction. The biggest successes here have been brought about by systems based on “deep learning” networks that build on the idea of the “perceptron” suggested 60 years ago by Frank Rosenblatt. However, achievements in the use of neural networks also uncovered a range of problems that cannot be solved using existing neural network methods.

First, any classic neural network model, whatever amount of data it is trained on and however precise it is in its predictions, is still a black box that does not provide any explanation of why a given decision was made, let alone disclose the structure and content of the knowledge it has acquired in the course of its training. This rules out the use of neural networks in contexts where explainability is required for legal or security reasons. For example, a decision to refuse a loan or to carry out a dangerous surgical procedure needs to be justified for legal purposes, and in the event that a neural network launches a missile at a civilian plane, the causes of this decision need to be identifiable if we want to correct it and prevent future occurrences.

Second, attempts to understand the nature of modern neural networks have demonstrated their weak ability to generalize. Neural networks remember isolated, often random, details of the samples they were exposed to during training and make decisions based on those details and not on a real general grasp of the object represented in the sample set. For instance, a neural network that was trained to recognize elephants and whales using sets of standard photos will see a stranded whale as an elephant and an elephant splashing around in the surf as a whale. Neural networks are good at remembering situations in similar contexts, but they lack the capacity to understand situations and cannot extrapolate the accumulated knowledge to situations in unusual settings.

Third, neural network models are random, fragmentary and opaque, which allows hackers to find ways of compromising applications based on these models by means of adversarial attacks. For example, a security system trained to identify people in a video stream can be confused when it sees a person in unusually colourful clothing. If this person is shoplifting, the system may not be able to distinguish them from shelves containing equally colourful items. While the brain structures underlying human vision are prone to so-called optical illusions, this problem acquires a more dramatic scale with modern neural networks: there are known cases where replacing an image with noise leads to the recognition of an object that is not there, or replacing one pixel in an image makes the network mistake the object for something else.

Fourth, the inadequacy of the information capacity and parameters of the neural network to the image of the world it is shown during training and operation can lead to the practical problem of catastrophic forgetting. This is seen when a system that had first been trained to identify situations in a set of contexts and then fine-tuned to recognize them in a new set of contexts may lose the ability to recognize them in the old set. For instance, a neural machine vision system initially trained to recognize pedestrians in an urban environment may be unable to identify dogs and cows in a rural setting, but additional training to recognize cows and dogs can make the model forget how to identify pedestrians, or start confusing them with small roadside trees.

Growth Potential?

The expert community sees a number of fundamental problems that need to be solved before a “general,” or “strong,” AI is possible. In particular, as demonstrated by the biggest annual AI conference held in Macao, “explainable AI” and “transfer learning” are simply necessary in some cases, such as defence, security, healthcare and finance. Many leading researchers also think that mastering these two areas will be the key to creating a “general,” or “strong,” AI.

Explainable AI allows for human beings (the user of the AI system) to understand the reasons why a system makes decisions and approve them if they are correct, or rework or fine-tune the system if they are not. This can be achieved by presenting data in an appropriate (explainable) manner or by using methods that allow this knowledge to be extracted with regard to specific precedents or the subject area as a whole. In a broader sense, explainable AI also refers to the capacity of a system to store, or at least present its knowledge in a human-understandable and human-verifiable form. The latter can be crucial when the cost of an error is too high for it only to be explainable post factum. And here we come to the possibility of extracting knowledge from the system, either to verify it or to feed it into another system.

Transfer learning is the possibility of transferring knowledge between different AI systems, as well as between man and machine so that the knowledge possessed by a human expert or accumulated by an individual system can be fed into a different system for use and fine-tuning. Theoretically speaking, this is necessary because the transfer of knowledge is only fundamentally possible when universal laws and rules can be abstracted from the system’s individual experience. Practically speaking, it is the prerequisite for making AI applications that will not learn by trial and error or through the use of a “training set,” but can be initialized with a base of expert-derived knowledge and rules – when the cost of an error is too high or when the training sample is too small.

How to Get the Best of Both Worlds?

There is currently no consensus on how to make an artificial general intelligence that is capable of solving the abovementioned problems or is based on technologies that could solve them.

One of the most promising approaches is probabilistic programming, which is a modern development of symbolic AI. In probabilistic programming, knowledge takes the form of algorithms and source, and target data is not represented by values of variables but by a probabilistic distribution of all possible values. Alexei Potapov, a leading Russian expert on artificial general intelligence, thinks that this area is now in a state that deep learning technology was in about ten years ago, so we can expect breakthroughs in the coming years.

Another promising “symbolic” area is Evgenii Vityaev’s semantic probabilistic modelling, which makes it possible to build explainable predictive models based on information represented as semantic networks with probabilistic inference based on Pyotr Anokhin’s theory of functional systems.

One of the most widely discussed ways to achieve this is through so-called neuro-symbolic integration – an attempt to get the best of both worlds by combining the learning capabilities of subsymbolic deep neural networks (which have already proven their worth) with the explainability of symbolic probabilistic modelling and programming (which hold significant promise). In addition to the technological considerations mentioned above, this area merits close attention from a cognitive psychology standpoint. As viewed by Daniel Kahneman, human thought can be construed as the interaction of two distinct but complementary systems: System 1 thinking is fast, unconscious, intuitive, unexplainable thinking, whereas System 2 thinking is slow, conscious, logical and explainable. System 1 provides for the effective performance of run-of-the-mill tasks and the recognition of familiar situations. In contrast, System 2 processes new information and makes sure we can adapt to new conditions by controlling and adapting the learning process of the first system. Systems of the first kind, as represented by neural networks, are already reaching Gartner’s so-called plateau of productivity in a variety of applications. But working applications based on systems of the second kind – not to mention hybrid neuro-symbolic systems which the most prominent industry players have only started to explore – have yet to be created.

This year, Russian researchers, entrepreneurs and government officials who are interested in developing artificial general intelligence have a unique opportunity to attend the first AGI-2020 international conference in St. Petersburg in late June 2020, where they can learn about all the latest developments in the field from the world’s leading experts.

From our partner RIAC

Anton Kolonin
Anton Kolonin
PhD in Technical Sciences, Founder, Aigents, Architect, SingularityNET, RIAC Expert