AjI: PERMANENT SELF-LEARNING

WHAT, WHY, AND HOW?

The ability to continuously and independently acquire knowledge is a natural trait of both humans and animals. Naturally, the requirement to implement such abilities should apply to any system that claims to possess intelligence.

The final stage in implementing such a capability is, of course, the program code. A natural problem in developing such code during the creation of General AI, Strong AI, or Artificial Just Intelligence (AjI) is that the requirement for this capability does not provide an obvious “roadmap” from requirements to source code. This chapter aims to fill this obvious gap. The collapse of naive expectations that LLMs and other neural network-based variants would spontaneously acquire this capability makes this aspect a key factor in AI progress.

From an AI developer’s perspective, when it comes to code, the first question is, of course, how should the knowledge that needs to be accumulated during self-learning be represented technically? This, in turn, depends on how the knowledge is intended to be used (in engineering terms, which operations need to be performed on it).

The minimum set of required methods for knowledge management is quite apparent:

• Knowledge is used for decision-making

• Knowledge is replenished (accumulated)

• Knowledge is verified for correctness

• Knowledge is modified, that is, some knowledge is removed and replaced with new knowledge without destroying the remaining knowledge

• Knowledge can be extracted for external dissemination (including explanations of decisions)

This set of capabilities is achievable through an approach that divides available knowledge into two subsets: facts and rules. The rules allow for finding new facts and constructing new rules by comparing/analyzing the accumulated facts. Facts and rules can have an external source – in particular, some are initially embedded in the intelligent system (their source is the system developers).

The requirement for self-learning obviously means that the intelligent system must be able to obtain facts from external sources (e.g., sensors, texts generated by humans or other systems). Less obvious is that the collection of facts for effective self-learning should be the goal of the system’s active, purposeful accumulation of facts, which makes the accumulated set of facts dependent on the decision-making rules.

An essential aspect of obtaining information from external sources in the form of text is that such information always represents a portion of knowledge already possessed by someone else (or something else). Receiving information in text form is a transfer of knowledge and cannot serve as confirmation of a system’s ability to learn autonomously. Developers of intelligent systems who design systems incapable of using any other ways of obtaining information besides texts risk falling into the same trap as the creators of LLM systems: the ability to correctly manipulate texts may turn out to be a facade for the lack of self-learning ability, reducing intelligent actions to searching for the required result in available texts, that is, in available knowledge created outside the system claiming to possess intelligence. Such a search may be complex – it may include logical analysis of information using rules that, having been constructed by someone else and obtained again in the form of text – but this is not veritable self-learning; searching among known knowledge does not create new facts or rules and therefore does not constitute the generation of new knowledge.

The circumstance described above may seem like a form of scholasticism at first glance. In reality, it reflects the unsuitability of an intellectual system that “feeds” exclusively on texts to perform tasks within the capabilities not only of preschool children but also of highly developed animals. Perhaps the main problem for the developer of an intellectual system is finding an area of ​​application for their creation, without which it is not viable.

Of course, self-learning involves not only expanding the set of known facts, but also expanding the set of rules that guide the system. Autonomy means that the rules can not only be received from the outside in a “ready-to-use” form, but also be generated by the system itself. This requires the existence of rules that enable the derivation of new rules from the available set of facts and rules; a minimal set of such meta-rules can and should be an initial part of the system.

Fact- and rule-based decision-making technology was actively developed in the early stages of AI research, including the creation of expert systems and machine translation systems that translated texts from one language to another. The enormous increase in the performance and information capacity of computer systems allows for a significant expansion in the number of rules and facts a particular system can operate on. A critical limitation of those early expert systems was their inability to independently replenish their fact base and to form new rules. These functions were assigned to users, meaning that expert systems could only assist humans, not act autonomously.

There are known variants for implementing AI’s ability to autonomously generate new rules (or modify existing ones) – primarily reinforcement learning and methods based on generating probability estimates of outcomes (usually a hybrid of these two approaches is used). What is surprising from our point of view is that, in most cases we are aware of, such developments differ radically from human practice in two respects: first, the experiments are conducted in extremely simple environments, and second, the self-learning does not make use of well-developed algorithms for experimentation and causal inference. In particular, a typical approach is to assume that the sought-after dependencies are probabilistic, whereas humans, both in everyday activities and in scientific experiments, assume dependencies are deterministic simply because they are.

In the experimental developments mentioned above, the simplicity of the environment significantly simplifies the task of symbolic (textual) description of the current situation and reduces the number of factors from which it is necessary to select those that actually belong to the rule (dependency); when using a real environment and sensors as a source of information, the situation is radically different, making it impossible to use algorithms suitable for a simple environment. Ultimately, the described approaches prove unsuitable as a prototype for a practically valuable system.

In summary, it can be assumed that success in developing Artificial just Intelligence (AjI) systems capable of self-learning can be expected in the following cases:

• Knowledge and facts are presented in a well-structured form (ontologies, semantic graphs, sets of rules, data tables)

• The constructed system is capable of extracting facts from the natural environment using sensors and actuators

• Individual facts and rules can be extracted, modified, and added selectively and permanently without disrupting others

• The initial (”innate”) set of rules includes rules for forming new rules

• The formation of new rules is capable of identifying deterministic dependencies and cause-and-effect relationships, without reducing the rules to probability estimation and without requiring the voluntary assignment of a degree of confidence in facts extracted in non-textual form.

• The emphasis in development is on the adequacy of the environment, not on the complexity and volume of knowledge.

The vast majority of the algorithms required for such implementations are known and tested, but a perception-concept gap still exists.