AI Hardware w/ Jim Keller:
In this video Jim Keller (CEO of Tenstorrent) talks about Tenstorrent's concept for their AI/ML hardware and compares it to GPU architectures etc.
At some point he starts to compare these concepts to the structure of the human brain and mentions "
cortical columns" (which we know from some of the Brainchip podcasts is also something PvM is interested to bring on the table for Akida's future). This is probably another concept out of the field of neuro science that therefore also found its way into the field of ML/AI and isn't anything exclusive to Brainchip. But nonetheless nice to hear other important players thinking in the same direction and maybe defining one of next Buzz-words
What I found even more interesting is another quote from Mr. Keller @32:27 regarding model training vs learning:
So however it will be implemented (hardware, software, spiking or else) one-shot and few-shot learning seems to become more important for the coming generations of ML/AI tech (at least according to to Mr. VanderMade and Mr. Keller).
What I'm still wondering or trying to understand is, if and how the self-learned knowledge inside of one or multiple hardware chips can be returned back to a more central instance/pool to actually get insights what and how the device actually learned to check for plausability etc.
But maybe I'm just trying to warp my head around too much things way above my pay grade ...