AGI: PERCEPTION-CONCEPT GAP

WHAT IS POSSIBLE AND WHAT IS NOT

As noted earlier AGI: PAST AND FUTURE - MISSING COMPONENTS, hardly one of the main obstacles to creating full-fledged AGI systems is the lack of such conversion algorithms of the flow of sensory information into a structured description of the situation, which would allow one to operate with previously unknown objects and, accordingly, provide the ability to generate new concepts. The lack of ability to detect unknown objects in the environment makes autonomous learning impossible. It is a source of dangerous failures in modern autonomous driving systems (as well as in other areas). With this chapter, we begin an analysis of possible ways to bridge the perception-concept gap; the analysis will be continued in subsequent chapters.

The modern AI/AGI development landscape includes two weakly interacting segments.

The first segment is the development of systems based on artificial neural networks (ANN) implemented by large companies; they consume enormous amounts of money and are widely covered (advertised) in the media. The latest achievements in this area are the so-called Large Language Models (LLM), including CharGPT and a series of analogs. The fundamental feature of this direction, in addition to being based on neural networks, is the dominance of an orientation towards the use of systems in web services mode, which, on the one hand, corresponds to the commercial goals of projects and, on the other is technologically determined by the need to use the most powerful computer systems that are unacceptable for the end user (including control systems autonomous robots, car autopilots, etc.).

The second segment is the development of full-featured AGI systems conducted by academic institutions and small companies using an alternative approach: semantic graphs, formal logic, rule-based inference, the absence of an expensive "training" phase before operation, and the explainability of solutions generated by the system. The amount of funding for these developments is thousands of times less. Accordingly, they are poorly represented in the mass media and are known mainly in a narrow circle of specialists.

A bridge between these two segments is the development of automotive autonomous driving systems. The funding here varies widely; logical components are used in addition to neural networks and local computing power (along with centralized ones at the "training" stage). This development group is the least likely to publish solutions' details.

The two segments of AI/AGI development have in common the gradual awareness by developers of the presence of the perception-concept gap discussed here. Naturally, instead of the "perception-concept gap," various terms are used since there is no established terminology yet, but the meaning remains unchanged.

As usual, the presence of a problem means a search for solutions.

One of the proposed "solutions" is formulated as "let's use a hybrid approach" - we will use both neural networks and technologies of alternative approaches. The idea of a hybrid in itself is reasonable. Still, firstly, in such a general formulation, the usable idea needs to be more specific. Secondly, hybridization is a double-edged sword: you can end up with a system that combines the disadvantages of both approaches.

The search for information about what methods of eliminating the perception-concept gap are analyzed/implemented/planned by developers led us to three publicly voiced options, which we will list in this chapter and explore in subsequent chapters. All three involve the use of approaches that are already in practice and have proven useful. If you know any intention to use options other than those listed below and they are not secret, we would appreciate receiving the relevant information.

In a brief formulation, the options sound like this:

1. Clustering data from sensors. The resulting clusters correspond to a set of concepts.

2. Detecting patterns in sensory data. New pattern - new concept.

3. Finding features and/or segmentation (as is the case in modern neural networks and/or computer vision); objects and a situation description are then constructed, which may include previously unknown objects.

An approach based on the hypothesis that as the size of the neural network increases, new abilities may arise that eliminate the problem under discussion is not included in our list due to its complete logical groundlessness.

In subsequent chapters, we will discuss the listed options concerning AI/AGI systems designed to work in the natural environment: the three-dimensional physical macroworld. Approaches, problems, and solutions for systems operating in virtual environments (chess programs, Go, computer games), dealing with an environment that can be described entirely formally at the stage of system development and lacking actual sensors, are fundamentally different from systems in a natural environment; the complexity of virtual environments is radically lower, which allows the use of much simpler approaches.