In this section, I've included description of my theoretical contributions to the field of artificial intelligence (AI). Please use the links on the left and colored links in the main text to navigate.
The most important aspect of artificial intelligence (AI) is machine learning. It has been the primary domain of my scientific interests. How are natural, biological mechanisms of learning constructed? How can learning be described mathematically on the most general level and in the most universal way? Efforts to answer these and related questions are intertwined. Novel information about natural mechanisms of learning prompts questions about how they could be described mathematically. New insights into mathematics of learning stimulate attempts to find out whether the related theoretical mechanisms are implemented in biological organisms.
The main goal of this project was to get insights into development of sensorimotor skills, which, for example, are required to perform every-day goal-directed movements such as reaching to specific target objects. Adding a novel feature allowed to create an AI system capable of learning to control arm movements very quickly from scratch.
It has been noticed a long time ago in the field of AI that sensorimotor skills require much more computational resources than the skill of cognitive reasoning (Moravec’s paradox). Therefore, it’s important to understand how sensorimotor skills, for instance, performing goal-directed arm movements, can be developed from scratch.
To be capable of learning those skills, a dedicated system has to first learn the dynamics of the arm as a mechanical object to form an internal model (IM) of that dynamics. I included several adaptive ANN components for implementing mapping-type functions, such as computation of control actions from the state of the arm, components of which are the arm joint angles and the corresponding angular velocities, and the desired, target state. Another set of functions is needed to represent prediction of the arm’s state resulting from a given vector of control actions applied at a specific state of the arm.
In comparison with the reinforcement learning technology (which is very general), it’s appeared possible to increase the efficiency of learning substantially by adding a novel functional component to the IM, namely, an ANN-based component that stores the best (minimal) cost of moving between two states. I call his component an IM of optimal path. Utilization of this functional component allowed the learning system to develop the skill in just a few trials (Shimansky, 2009). The control system was learning online, as the data were collected during an ongoing goal-directed, controlled movement. Additional innovations, such as usage of data pattern cashes, allowed the learning system to function very efficiently, avoiding a necessity to process massive amounts of training data and updating the ANN components concurrently with making movement control trials.
The IM of optimal path between two states is a generalization of storing the best-so-far found solution of a simpler optimization problem, where, e.g., a function of several arguments requires minimization: standard algorithms of incremental minimization include the least value of the objective function (together with the best vector of argument values found so far), so that the least function value could be compared with other values and, if a better one is found, the old values are replaced with the new, better one.
Finding an efficient biologically plausible method for training ANN has been a long-standing problem. I solved it by using a method very similar to what bacteria use for chemotaxis.
To train artificial neural networks (ANN) for so-called deep learning, the method of error backpropagation (EBP) is a current standard, often in combination with simulated annealing (to escape shallow local minima). As well known, EBP, besides being biologically implausible (it is not at all clear how it could be implemented in a biological organism), has considerable limitations in regard to the training procedure (too rigid) and ANN topology (it is not friendly to loops in the ANN graph). Therefore, it is rather desirable to develop a method for training ANN that would not have those drawbacks without losing efficiency. It appeared possible to develop such a method based on kinetic force principle (KFP Shimansky, 2009) I discovered earlier. This method, regulation of modification probability (RMP) turned out even more efficient than EBP (Shimansky, 2009). Interestingly, RMP can be used to model chemotaxis in bacteria (Shimansky, 2009).
There are several advantages of using RMP technology of training ANN.
There is a fundamental problem with computer-based implementation of artificial self-learning systems: in general, learning cannot be described by an algorithm. However, there are strong reasons to believe that learning can be physically implemented, thus transcending the limits imposed by computability requirements.
Theoretically, learning is essentially a search of optimality, which is based on enumeration of a certain set of possible variants, and there are important sets that are not algorithmically enumerable (most notable, the set of all fully defined functions of integer arguments). One way to implement such enumeration is via solving the so-called halting problem, which does not appear to be physically achievable. At the same time, however, a theoretical analysis showed that universal learning systems (capable of learning to control optimally any object from a given class in finite time), must be physically possible even for algorithmically non-enumerable classes (Shimansky, 2004). Looking for practical utilizations of that possibility led me to the concept of trans-algorithmicity (Shimansky, 2018).
Algorithms cannot change themselves. Due to this limitation, intelligent human individuals with computer programming skills are required to alter algorithms encoded in computer memory. Human mind, unlike machine, is capable of moving in the space of algorithms. In other words, human mind is trans-algorithmic, which allows it to be creative and learn what machine cannot learn by itself.
A principle limitation of computer-based learning systems is that any algorithm cannot modify itself (this is why a team of programmers and engineers with natural intelligence is required to adjust software and/or hardware in response to external changes). How is it possible to design a learning system capable of optimizing search in the space of different algorithms? Besides being an interesting and challenging theoretical problem (related to the nature of biological organisms, whose capacity to learn and adapt do not seem to be limited by any fixed algorithm), a solution of this problem would have valuable practical applications, especially in unmanned space expeditions to remote planets, where the possibilities of manual control from Earth are rather limited. Based on the concepts of ULS and KFP, I developed the concept of trans-algorithmicity (TA), which, in brief, is a capacity to move in the space of algorithms. Although TA obviously cannot be based on any fixed algorithm, there are strong reasons to believe that it is possible to implement in physical systems. The main idea behind TA is that TA system opens itself to environment-produced modifications (which are generally available due the ubiquity of fluctuations in the universe) in the system’s physical design and regulates the intensity of such modifications to improve itself with respect to an objective function describing its vitality (in other words, the criterion of its optimality) (Shimansky, 2018).