Importance of a Modular HTM Layer



We talk a lot about how HTM algorithms work within one layer of cortex, but new theory introduces multiple layers. This is an important concept to understand above and beyond understanding SDRs, SP, and TM.

As a software engineer, the “modular” aspect of the cortex’s Column/Layer structure is very appealing, because we can think of a layer generically as a compute module. It has input and output and it computes over time. The internal workings of the layer might be a black box from the outside. But as long as we know what it is doing and what it’s input and output represent, we can still use it.

There are many layers of a Cortical Column, and it is clear that each one is operating a bit differently from others. By treating Layers as modular software components, we see how to build compute circuits out of these modules. The sensorimotor circuit we talk about is a cortical circuit made up of two compute layers. Ongoing research depends on even more layers of the Cortical Column getting filled in. Things seem to be moving quickly!

In the meantime, let’s not forget about this modular structure that will be the framework of future intelligent architectures. It is probably not so often that a biological structure maps so nicely into a software component. As you build your own HTM systems, keep these structures in mind. Think about the ways dimensionality in the data is transformed at each step, how representations are mixed, merged, and unrolled.

It is an amazing time to be studying such an intriguing topic! :smile:

Object / concept representation in high-dimensional space

5 posts were split to a new topic: Object / concept representation in high-dimensional space

Object / concept representation in high-dimensional space

This is precisely the thinking I had in mind when I looked at designing HTM.Java’s Network API. Modeling Layers as “Observables” which can be transformed (used as many typed operations such as arithmetic, temporal (delayed & cumulative), boolean logical, and many other combinations) - all with streaming and idempotent isolation of each “unit” in the “functional” way. :slight_smile:

@rhyolight I’ve been “away” for a bit, has there in fact been more inclusion of other layers in the current theory?


Our next paper on location will include models of two more layers in another circuit.