The brain does not just process the information – it tries to find its meaning. The ability to find the meaning of the information is one of the most important features of our brain.
We believe that all information is inherently polysemantic. The same information may be interpreted differently in different contexts. We think that the only way to find the meaning of the information is to look through all possible contexts, interpret the information in each of them, and determine which interpretations make sense.
In order to check if an interpretation is viable, we need to test it by comparing with previous experience. It looks like our memory is designed in a way that allows it to perform computation in place and recognize a familiar interpretation.
The cerebral cortex is organized in regular structures known as cortical minicolumns. Each minicolumn contains about a hundred neurons. Each cortex area contains about a million minicolumns. The amount of minicolumns in the whole cortex is close to a hundred million. We have strong reasons to believe, that each cortical minicolumn acts as an independent computing module. Each minicolumn is associated with its unique context. A context is a set of rules that define, how the input information should be transformed in order to get the interpretation specific to this context. In a sense, each minicolumn sees and estimates what’s happening from its own “point of view.”
In our research we state, that memory is not the strengthening or weakening of synaptic connections. Instead, we propose the memory mechanism that is based on the specific interference of information-coding patterns of neural activity. In our theory, the origins of memory are the clusters of receptors scattered across the membranes of neurons. At the same time, these clusters perform the in-memory computations.
It has long been known that memory is not localized in the brain somewhere specifically. However, the underlying mechanism of its distributed storage remained a mystery. We believe, that each minicolumn does not hold just a fragment of shared memory. Instead, along with other minicolumns, it contains a full copy of all the memories, specific to this cortex area. This is needed to allow each minicolumn to independently estimate the validity of its interpretation. We were able to show the biological validity of such an assumption.
We have shown that to solve combinatorially difficult problems one can use the method of splitting a high-dimensional initial problem space into a large number of overlapping random subspaces of a much smaller dimension. Solutions that are not accessible to conventional iterative methods in the original space turn out to be obvious in some “successful” subspaces. With the right configuration of the combinatorial space, we may be confident that the answer will be found eventually. We have shown that this technique may be used by the cortical minicolumns to perform computations. It is equally applicable to build the context space on top of a cortex area.