I was talking about HTM theory to one of my colleagues. The point he made was about the encoder in the HTM that converts different stimuli/sensory information into a single type of data/SDR.
My colleague has a biology/chemistry background.
His understanding is that the brain is a complex unit
where signals and stimuli are transmitted among countless chemicals, electrical, and physical routes and are stored and processed in equally diverse mechanisms and that he hasn’t seen any evidence of an encoder.
Has there been any neuroscience discoveries regarding an that - an encoder in our neocortex that does this kind of conversion for different stimuli into a single
From the different papers and that Jeff has written, I’ve learned that the neocortex is divided into Corticol columns(discovery we made fifty years ago) and there’s a common algorithm/mechanism to each of those Corticol column works
As a follow up question , do we also have any neuroscience papers that support this idea as well - ’
“common algorithm/mechanism to each Corticol column”? I know that Mountcastle discovered the structure of the Corticol columns 50 fifty years ago but I haven’t found any papers that talked about this common algorithm.
HTM combines top-down and bottom-up approaches to neuroscience. It draws on specific observations, but the goal is to figure out what the cortex does, not exactly how it does it. Different parts of the neocortex might accomplish different things differently, like forming SDRs, and they certainly have specializations. Still, the neocortex must use a common algorithm because it has a mostly homogeneous circuit and because regions for one sense can be re-allocated to other senses. There might be some things necessary for intelligence which every region does differently, but those will be implied by the rest.
HTM doesn’t exclude multiple types of data or parallel sensory streams. You could add things like rate coding or sense-specific things to SDRs, while still retaining the benefits of sparse representation. Those benefits don’t just imply optimizations. For example, being able to represent multiple things with the same population of neurons allows things like HTM’s temporal memory.
The encoder is sort of like the sensory organ. It’s not meant to be biologically accurate. It just produces a suitable input to the rest of the system. SDRs are formed in the next processing step after the encoder, which is the spatial pooler. The spatial pooler is like minicolumns in the neocortex, not to be confused with macrocolumns a.k.a. cortical columns. We don’t know exactly how it works, and there probably isn’t a strictly universal algorithm, but the end result is that columns of neurons have similar sensory response properties.
Activity in the neocortex is sparse, and it’s distributed because each neuron is connected with thousands of others so a single excitatory neuron has little influence on the others.
HTM doesn’t actually claim the cortical sheet is split into these discrete units. Sometimes the cortex really is, but maybe not always. It’s just easier to think about a discrete patch of the neocortex. Each neuron mostly interacts with other neurons in a radius of ~.2 to several millimeters, depending on the species and cortical region.
The common algorithm of the neocortex is basically the connectivity and neural properties, and there are lots of papers about those. The algorithm is very difficult to infer because every cortical region and study is different, and because there might be differences in implementation. I imagine one region evolves from another, and then mostly keeps the same algorithm but evolves some specializations, especially for recognizing basic sensory features like oriented lines. There might even be some holes in the algorithm which each region fills in on its own.
The cortex is complicated, so there are lots of variations between regions, but each region is far more similar than different. The similarities are important things, like having 6 layers, certain major classes of pyramidal neurons in each layer, and certain major connections between each of those. That’s surprising considering the cortex has been around for a couple hundred million years.
I think it’s not a complete answer. The encoder typically combines several types of input, which are treated by SDR as a unified set of elements, as I undersend, that’s the core of the question.
But I don’t see why it would provoke any concerns from a person with a biology/chemistry background. You shouldn’t look for such encoder as a dedicated structure, the brain is highly interconnected structure, where one module (represented in this case as SDR) usually receives input from many other modules.