Stephen Grossberg’s ART (Adaptive Resonance Theory) is a circuit that clusters inputs into concepts, and when inputs are close enough to a prototype, they can modify the weights to that prototype, but if they are far enough away from it, they can’t modify it, and an automatic search is done for a category that matches the weights better. If none is found, then a new category is learned. This means some generalization is done, but only within limits. I was wondering what controls generalization on dendrites and of concepts in Numenta’s theory.
I think Dr Grossberg sometimes participates in the forum - he noted he had published a book Conscious Mind, Resonant Brain, about his work. ART as he outlines it there aims to avoid the stability-plasticity dilemma, which to complexity theorists (me anyway) appears to equate to the edge of chaos between chaos and stability. He argues It is possible to avoid total stability and chaos by using resonance to generate learning. He has many examples from different types of brain processing of sensory inputs and shows how these can be explained theoretically and with model circuits that operate over laminar cortical networks. To me this appears to be be the point of alignment with HTM although he may disagree or clarify.
Within the logic of ART Grossberg develops a series of mathematical formalisms that describe the processes performed by complex networks of neurons. In particular the ART matching rule appears to provide the mechanism Mark Springer is describing - I believe - when the circuits are being applied to object matching and categorization.
Hope that helps until Dr. Grossberg clarifies and Numenta comment.
This thread should help ://discourse.numenta.org/t/conscious-mind-resonant-brain-how-each-brain-makes-a-mind/8754
Grossberg’s work uses a dynamical systems approach. The work is much broader in application and scope than HTM. For example Grossbergian systems have been realized for object recognition in 3D scenes using binocular vision with saccades.
The HTM neuron maps to a discrete binary implementation well. Grossbergian systems can be modelled with digital computers but are also modelled with analog systems and spiking neural networks.