Some years ago I studied deep learning. There were good books, several frameworks and plenty of trained neural networks for free. In other words, there is interest that people learn to use deep learning.
On the other hand, HTM which seems much powerful than deep learning doesn’t have an updated framework with proper working examples.
Besides, the Corticol.io’s white paper is a joke. It’s just an ad.
So, imho, numenta is interested in creating companies that use HTM but doesn’t want peopple and researchers to use it.
I was very excited about HTM but now I am very disappointed with the lack of resources to develop good AI tools.
There’s a thread that explains the reason that NuPIC was put into maintenance mode back in 2017: NuPIC 1.0 has been released
HTM clearly has massive potential given the capabilities of human intelligence, but there’s more research ahead and that’s where Numenta’s effort is focused.
Nupic was open source and so after the old maintainers of it (Numenta) lost interest the “community-fork” of nupic started up. We got together and made a lot of long needed improvements, and then rebranded it as “htm.core” so that people would not confuse the two (now quite different) code bases.
There are also several other good HTM implementations, which have been posted on this forum from time to time.
I believe a lot of it has to do with the name “hierarchical temporal memory” its kindof a mouthful and the opposite of catchy, I dont know why neuroscientists love so much the word “hierarchical”.
I’m also sorta bad at naming things, but I’d have called it something else, maybe “wide learning” or maybe “cortical storage”
Let’s not be naive, While Numenta does share a lot of its findings, in the end, it is a commercial operation. While they provide the basic framework to develop applications, try not to think about them in terms of ( Pytorch or TF ) where things are spoon-fed.
To your point about Cortical IO, because they have not opened up their framework and it is not subject to outside scrutiny or explainability, no one in the broader deep learning community takes them seriously. Even if is a secret, if there was any value in what they are doing they would have bought over by the likes of Google or Facebook.
My two cents take what you can and see if applies to your domain. If you got some good ideas to improve or enhance please contribute.
I think what might be very useful for the htm.core codebase would be an examples directory where working implementations (chatbots, robot control, data analysis, etc.) could be placed.