has anyone taken a good look at their tech as being more efficient?

I think has a far more efficient tech as to machine intel. Pros? Cons?


I looked into them a few days ago, but I could not get to their code, and I don’t quite understand what they are doing. They say their code is open source, but it’s not really because you have to fill out a form and join a secret club to get to it. I don’t understand that.

Has anyone actually joined and seen the code?

1 Like

It’s here, but it got a proprietary license:
background paper is here:

From what I can gather (in a few minutes of research) they want to implement part of the algorithm in a memristor array.
I believe the HTM algorithm is a good fit for memristor arrays as well. It’s kind of like an analog ‘gpu’ for summing values from many inputs (synapse / dendrites). I expect to see HTM running on a memristor array or something like it within the next 10 yrs - it’s one of the few ways to scale up, as GPU is problematic. I’m sure Jeff has some thoughts on this, and porting HTM to neuromorphic silicon in general.


OK, I’m still researching Knowm - I went down the rabbit hole and it’s pretty interesting.

The software is pretty damn good at temporal prediction and what may be online learning - not totally sure on that count, needs more research. Given a complex wave (multiple sine waves combined) it was able to track it accurately after a few thousand iterations. The code to generate the wave is there and could be used to compare HTM results on the same test. Results here: It would be intersting to challenge them to enter the Numenta Anomaly Benchmark Competition.

They have some sort of memristor hardware they call kT-RAM and I believe they simulate it in software at this time. So whatever software they are running simulates this kT-RAM that they have patents on and supposedly are building. From what I can gather it is a general purpose neuromorphic memristor crossbar storage device designed to emulate synaptic connections. It’s worth taking a closer look to see if it might be usable in some future version of HTM to store all the connections in the columns/synapse/dendrites

At any rate the tech and ideas behind it look promising, but the materials as presented on the web site are a bit hand-wavy and at first I was thinking this is sketchy. You would need to dig into the code and patents to see what is behind it. When I did I was shocked to see several patents that were granted in 2011 for HTM implementations in silicon. This guy, Alex Nugent has been thinking deeply about how to build neuromorphic silicon that can break the von Nuumann bottleneck that is currently stopping HTM from scaling. I think folks here should take a closer look at what is going on at Knowm and the ideas behind crossbar memristors and neuromorphic techniques that can scale HTM. Alex Nugent / founder info

HTM impementation patents: