Appreciating that not everyone tries to optimise for LLMs and we are still doing things like this. If you're looking at HN alone, it sometimes feels like the hype could drown out everything else.
It's just a single linear layer and it's not clear to me that the technology is capable of anything more. If I'm reading it correctly it sounds like running the model forward couldn't even use the technology, they had to record the weights and do it the old fashion way.
Nice that they can do the processing in the GHz range, but from some pictures in the paper, it seems the system has only 60 'cells', which is rather low compared to the number of cells found in brains of animals that display complex behavior. To me it seems this is an optimization in the wrong dimension.
Maybe try simulating the algorithms in software before building hardware? People have been trying to get spiking networks to work for several decades now, with zero success. If it does not work in software, it's not going to work in hardware.
>If it does not work in software, it's not going to work in hardware.
Aren't there limits to what can be simulated in software? Analog systems dealing with infinite precision, and having large numbers of connections between neurons is bound to hit the von Neumann bottleneck for classical computers where memory and compute are separate?
“Zero success” seems a bit strong. People have been able to get 96% accuracy on MINST digits on their local machine.
https://norse.github.io/notebooks/mnist_classifiers.html
I think it may be more accurate to say “1970s level neural net performance”. The evidence suggests it is a nascent field of research.
Ghz speed video processing, even if we only get very basic segmentation or recognition out of it, is probably crazy useful. Need to face recognize every seat at a stadium?
Well, if you have enough cameras, 60,000 seats could be scanned 250 thousand times a second. Or if you want to scan a second of video at 60fps, you'd be able to check all of them at a mere 4 thousand times a second.
Anyway, good to see interesting raw research. I imagine there are a number of military and security use cases here that could fund something to market (at least a small initial market).
10 comments
[ 2.4 ms ] story [ 20.8 ms ] threadAren't there limits to what can be simulated in software? Analog systems dealing with infinite precision, and having large numbers of connections between neurons is bound to hit the von Neumann bottleneck for classical computers where memory and compute are separate?
Well, if you have enough cameras, 60,000 seats could be scanned 250 thousand times a second. Or if you want to scan a second of video at 60fps, you'd be able to check all of them at a mere 4 thousand times a second.
Anyway, good to see interesting raw research. I imagine there are a number of military and security use cases here that could fund something to market (at least a small initial market).