#neuromorphic #ComputerVision #EventBasedVision

My humble hope: this could be a turning point for SNNs to excel in what they were designed for: sparse, spatio-temporal signal processing.

The best part? Everything is open-source. Steal it, modify it, send it to hardware with the Neuromorphic Intermediate Representation - just please cite us :-)

Can you unpack this a bit?
Some argue that large models work well in machine learning because of the mysterious fact that gradient descent improves at scale, despite non-convexity (arxiv.org/pdf/2105.04026).
Would you agree? If so, how does this apply to simulations?

Ah, yes, thank you. I initially read the quote to mean that physics restrict the algorithm, not that physics IS the algorithm.
For finding solutions, as you write, this distinction is important. Restrictions have to be baked in from the beginning, otherwise any “solution” will be meaningless.

This is actually interesting. Did she believe that the role of silicon in VLSI systems is similar to the role of neural substrates in nervous systems?

If so, I would agree with Brad that I don't see the big difference. But simulations will always be a poor man's approximation

Why stop there? If I had something to sell, I would want to hijack every neuromodulator I could get my hands on. Eternal chemical bliss 🏴‍☠️

That's a great point! We cannot equate the hardware with the model. "NeuroAI" is indeed not a model.

I wonder whether the ambiguity would stand if we had a solid understanding of how the substrate related to the algorithm. Where does physics/hardware stop and where does computation begin?

Oh dear, that's terrible and borderline denigrative 😬

I'm wondering how to address this. Isn't part of the reason why some words remain less viscous that they have strong definitions? Could it be that part of the problem is that #NeuroAI is too vague? What if we need better definitions?
We could start with "intelligence"...

I agree that languages inevitably evolve, but at the same time words have to *mean* something.
Personally, I consider "neuromorphic" to apply to concepts outside hardware. I am open to changing my mind, but there have been so many conflicting takes on this that I am, frankly, confused.

It seems like a neat paper on DSP, but could you tell me how this relates to continous computation?

I think that's exactly the right mindset. It'll be hard to balance concerns when the new wave of hardware hits, but sticking to the "fast weights" bit is crucial. Nice.
My hunch still is that this requires a continuous representation, but I may be wrong 🤔 maybe we should do a survey?

open-neuromorphic.org ☺️

I'm still not sold on the MLIR angle. It may help integration of existing models, but MLIR is inherently digital. Wouldn't that hinder the computational expressivity of mixed-signal hardware?