Necessary and sufficient conditions for the cortex-like representation

This is a development of Synthetic Intelligence from Zero Principles from the perspective of the required for SI representation.

1. Modality agnosticism

As was shown by Mountcastle, even though each modality needs specific encoder for the most efficient processing, the neocortex has a similar structure for all sensory modalities and even abstract ideas. The same structure means the same computation, therefore the same representation of the data.

One of the aspects is the same representation for temporal and spatial patterns.

2. Intrinsic invariance

The representation has to be immune to any possible change. Such invariance doesn’t need multiple variations of the same pattern for learning, what is demonstrated by mammal learning.
Intrinsic invariance can’t be replaced by an inner transformation of a pattern because of its compositionality: it would need separation of all possible parts (and part of parts) and repeating all transformations for each of it.

3. Implicit soft compositionality and hierarchy

Any part of the pattern can be or can not be regarded as a compositional part of the general pattern, but has to be present implicitly. For example, whiskers and ears are parts of a cat, but, by adding everything else from the same level of the hierarchy, you wouldn’t receive the whole cat. Also, there isn’t an obvious border between the cat’s ear and everything else. And some parts can be distinguished or not (like cat’s cheeks — is it a thing?). So, compositionality is not hard or strict and real patterns have a weak hierarchy.

At the same time, people demonstrate the capability to describe several different hierarchies and ways of decomposition for the same object, what needs inference at many levels of it. For instance, a cat can be represented as a set of body parts, or as front-back (left-right, bottom-up, inner-outer) halves, or in terms of biological classification (animal >> vertebrates >> mammal >> cat family >> Siamese cat >> …), or as a set of behavior features, or sets of organs/tissue, etc.

4. Intrinsic multidimensionality

The representation has to have the capability to merge (but not dilute) absolutely different dimensions of a pattern.

For instance, any part of a sentence is an intersection of semantics, syntax, morphology, etc. and any deviation from the norm exists in the context of a specific dimension. So, independent structure for each dimension has to be supported for analyses, but such analyses can’t be fully carried out without the context of other dimensions.

Or, in the case of looking at an open book, we perceive a point which has a color, saturation, intensity, can be regarded as a part of rectangular/ symbol/word/title/page/book/desk/room, it is soft being regarded as a part of the paper and hard as a part of the book, etc. All these dimensions exist simultaneously and can be or can be not important in the current context.

5. Ability to represent the world structure and integrate modalities through it

In our mind we have a virtual model of the word which is a basis which allows us to efficiently orient oneself in the real world, and even the most abstract ideas exist as projections to it.