Historic failure at generalization
Perhaps AI’s greatest failure in its 70-plus-year history has been its inability to realize human-like generalization in its machine of choice. The failure actually may be fundamental to AI’s main theory of mind, computationalism, and seems to be at the heart of theoretical dilemmas including the frame problem (McCarthy and Hayes, 1969), problem of common-sense knowledge (Hubert Dreyfus, 1965) and the problem of combinatorial explosion (James Lighthill, 1973).
Principles
According to renowned AI researcher, theoretician and aficionado Ben Goertzel in his recent article, From Here to Human-Level Artificial General Intelligence in Four (Not All That) Simple Steps,
…we don’t yet know enough useful theoretical laws of general intelligence to proceed with AGI engineering in a mainly theory-guided way;
More accurately, AI doesn’t know any laws, useful or not, of general intelligence. So AI doesn’t know the principles of generalization. What are the principles of generalization?
Repetition
The idea is that the most fundamental principle is repetition. That which is repeated the most is the most general.
Example
The principle is applied from the beginning of building a semantic structure. Serial input is assumed. The input is one sensory stream. This “contains”: (a) objects, (b) instances of the relation between objects of temporal contiguity, and (c) a repetition count of instances of the relation.
The count is incremented by one when a pair of objects arrives that has the same values of the same property (usually in the same order) as the base instance of the count.
Using linguistic symbols and values of the property of shape to illustrate this, if at the start, A follows B then this pair of objects is stored physically together or otherwise associated in a temporary storage area and given a count of 1. Each further time an object whose shape is A follows in the stream an object whose shape is B, the further relationship instance is not stored but rather the count of the base instance is incremented.
(The processing machine is so constructed that it reacts to slightly different shapes as though they have the same shape, i.e., the same reaction occurs.)
If in a processing session only the symbol pair of the greatest repetition count is stored, and if A follows B has the greatest repetition count then the pair B then A (symbolized “(BA)”) is stored, taking the first to arrive as being on the left.
If in the next session, A follows B then this pair is not stored but rather the stored instance is flagged as activated, and if C then follows (BA) in the stream and (BA)C has the highest repetition count in the stream during that processing session, then C is permanently stored and connected to the node that represents (BA), hence now ((BA)C) is stored, which is less general – more detailed – than (BA).
associative-ai.com