On-chip Face Recognition System Design with Memristive Hierarchical Temporal Memory from Nazarbayev University in Kazakhstan.
Wow very ambitious.
Speaking of HTM in hardware:
I’ve been waiting for someone to do this.
Writing encoders and decoders as functions in Haskell could be a fun start. The result would be a nice and compact module.
Let’s wait for Numenta to put the “machine” back into "machine intelligence"
before putting effort into implementations.
It’s a strange paper: they propose a novel face-recognition system, but completely fail to justify it - why would someone want to build it using HTM, and not a good old convnet? Power efficiency? Accuracy? Chip area? Of course, none of those will be known until the chip is built and tested, and it’s not even clear if they themselves plan to ever build a prototype.
Memristors are really interesting. They are a fundamental electronic paradigm - up there with transistors, resistors and capacitors. They are a fundamental building block.
They have some really interesting properties that map to maintaining some kind of “memory” based on previous electrical inputs. Does that ring a bell? Sound familiar?
If you have a few minutes: https://youtu.be/bKGhvKyjgLY?t=168
Memristors allow you to model a weighted input with the ability to change the weighting and doing it without power at (in the future) high densities – millions or billions per mm2. That’s the hardware for the HTM software.