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.
There is massive hype, no doubt about it, but lets also not forget how LLMs have basically solved NLP, are a step change in many dimensions and are disrupting and changing things like software engineering like nothing else before it.
So I hear you, but on the flip side we _should_ be reading a lot about LLMs here, as they have a direct impact on the work that most of us do.
That said, seeing other papers pop up that are not related to transformer based networks is appreciated.
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).
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.
Is this actually capable of propagating the gradient and training more complex layers though?
A lot of these novel AI accelerators run into problems like that because they're not capable of general purpose computing. A good example of that are the boltzman machines on Dwave's stuff. Yeah it can do that but it can only do that because the machine is only capable of doing QUBO.
For inference we do not care about training, right?
But if we could make cheaper inference machines available, everyone would profit.
Isn't it that LLMs use more energy in inference than training these days?
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.
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.
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.
There is massive hype, no doubt about it, but lets also not forget how LLMs have basically solved NLP, are a step change in many dimensions and are disrupting and changing things like software engineering like nothing else before it.
So I hear you, but on the flip side we _should_ be reading a lot about LLMs here, as they have a direct impact on the work that most of us do.
That said, seeing other papers pop up that are not related to transformer based networks is appreciated.
Thank you, brother. Besides not all that goes in HN is strictly LLM, really dunno why the scare.
I couldnt agree more.
Retina-inspired video recognition using light. Cool. May be a visual cortex next year.
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).
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.
Would you have discredited early AI work because they could only train and compute a couple of weights?
This is about first prototypes and scaling is often easier than the basic principle.
Is this actually capable of propagating the gradient and training more complex layers though?
A lot of these novel AI accelerators run into problems like that because they're not capable of general purpose computing. A good example of that are the boltzman machines on Dwave's stuff. Yeah it can do that but it can only do that because the machine is only capable of doing QUBO.
For inference we do not care about training, right?
But if we could make cheaper inference machines available, everyone would profit. Isn't it that LLMs use more energy in inference than training these days?
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?
It's not clear that "infinite precision" is a meaningful thing. All inputs and outputs even to analog systems will only ever have finite precision.
“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.
This seems to work in hardware, per the paper. At least to 80% accuracy.
A lot of unrigorous claims for an abstract…
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.
I suspect practicality is not the goal here, but rather a proof of concept. Perhaps they saw speed as an important technical barrier to cross