Hello, I am very interested in HTM theory and have read relevant papers. I think HTM theory is very cool. Since I am not particularly familiar with neuroscience, my question is, which theories of HTM have been confirmed by neuroscientists so far? Can you provide relevant papers? Thank you very much!
It is the other way around - HTM is closely based on the biology.
Thanks for joining @Zoey_Lee. Mark is right, HTM is very closely based on common accepted neuroscience. If you read our papers you will see that they cite relevant neuroscience papers that helped us develop our theory.
I would caution people just saying HTM is supported by ‘x’ papers. As for @Zoey_Lee question, it’s neural plausible, not necessarily neurally confirmed.
If you don’t want to read every one of the cited papers, I think the main take away from neuroscience and cortical columns that HTM has clung to as its core concept is mostly just the SDR aspect of it. Everything else seems to be made to enforce that sparseness. The idea being that if two codes overlap significantly enough, then they can be considered the same or similar enough. Then essentially what you are doing for sequences is finding the similarity between the current state PLUS the previous states compared to the viewed code.
As far as being closely related to biology, I would have to disagree. HTM theory does many things that the brain does not do, solely to get the same effect as the brain. Like the global and local inhibition functions. The idea is that an interneuron plays a role on deciding which of a cluster of neurons fires. The first that fires sends a signal to the interneuron and the internueron stops the others that are connected to it from firing. What HTM does is replicates that sparse functionality by having inhibition radiuses or even a global modifier that just lets the top x% fire and all of the others don’t.
So I wouldn’t say it is closely based on biology, but it has taken lessons to try to replicate similar results that are important in what makes intelligence. But in very different ways.
I don’t quite agree with your statement. I’ve always said that HTM was closed based on biology, and I think it is true. You are right that HTM is not a complete similation of neurons in the neocortex. If you want that, there are other scientific endeavors attempting that simulation. We’ve never tried to to that, because we are trying to understand the core mechanisms of intelligence in the brain. We’ve always said that if any legitimate experimental evidence arises that contradicts HTM theory, we would have to account for it, even if it meant changing the theory to accommodate what is really happening in biology. Since 7 years I’ve worked here I have not seen that happen (although the research team has done this a lot internally before producing papers).
There are plenty of things that HTM is not aligned with in terms of biology. For example global inhibition I mentioned earlier. The only reasoning behind global and local inhibition that I have seen so far is that ‘there is inhibition so we inhibit in A way’. If that weren’t the case there would only be the one choice and it would be closely related to the reasoning and results of biology. Which that is completely fine, I’m not saying that you need to know everything to start working on a theory. That is by definition what a theory is, you don’t know everything and you are trying to work it out. But as scientists, if you start assuming “the way we are doing it is getting results so lets just assume that’s the way biology does it until someone tells us different” then honestly that’s only slightly different than the current AI community.
It maybe the case that things like inhibition method, biasing local spatial processing before temporal, ignoring minicolumn functionality for the overall hypercolumn results, resetting, etc., might not actually affect the overall effect that cortical columns create. My point is that there are things that have been done that exclude key components, like a neurons connection to higher and horizontal neighbors, that make the assumption will give the same results. A layers inhibition methodology is the replacement for lateral connections and interneurons. And as scientists, you can’t possibly comfortably say that that method its neurally confirmed rather than neural plausible.
That’s my point, really that it’s a step toward proving but design choices have made the theory much different than biology’s methods but still trying to stay true to the end result. Hopefully with no unintended consequences of biasing the learning in certain ways.
Also I agree with the whole, “t’s not a simulation and doesn’t need to be” sentiment. If you can just get either the minicolumn or hypercolumn functionality right, than aside from connecting other modules and their functionality of the brain, then you don’t need an individual neuron simulation.