Coffee BeeJul 16, 2024
Florian Marquardt

For friends of #neuromorphicComputing, we are happy that our new idea, spearheaded by Clara Wanjura, is now published in #NaturePhysics. It shows how to take a linear optical device and turn it into a fully nonlinear neuromorphic device which can be trained efficiently.

https://www.nature.com/articles/s41567-024-02534-9

Neuromorphic computing is all about possible paths towards replacing energy-hungry digital artificial neural networks by more physics-based devices.
...

Picture credit: CC Wanjura

1/n

Fully nonlinear neuromorphic computing with linear wave scattering - Nature Physics

As the energy consumption of neural networks continues to grow, different approaches to deep learning are needed. A neuromorphic method offering nonlinear computation based on linear wave scattering can be implemented using integrated photonics.

Nature
Jul 16, 2024 at 3:08pmWeb
Show threadFlorian MarquardtJul 16, 2024

Optical systems are a major contender for useful neuromorphic systems. Unfortunately, linear optical systems are usually not good enough: their "expressivity" is limited, because they can only represent a linear function of the input signal.

However, in our article, we show a way of how to circumvent this restriction. The main idea is that you can inject the input into the device not in the form of a light field, but in the form of tuneable parameters.

2/n

Show threadFlorian MarquardtJul 16, 2024

The scattering matrix of an optical system depends in a nonlinear way on parameters (such as tuneable resonance frequencies of optical resonators). In this way, a linear optical system can be turned into a nonlinear information processing machine.

The second major idea we introduce in our work is an efficient physics-based way of training such devices. Training neuromorphic systems is not that easy. In particular, scientists are looking for efficient physics-based training.
...

3/n

Show threadFlorian MarquardtJul 16, 2024

We show that in a linear optical system, physics-based training becomes quite easy: To find the dependence of the scattering matrix on the many tuneable parameters, one could naively do many scattering experiments, but that would be slow. We show a way to instead perform only a few scattering experiments, and deduce training gradients from those (in some physics version of backpropagation).

Our article: https://www.nature.com/articles/s41567-024-02534-9

News & Views by Peter McMahon: https://www.nature.com/articles/s41567-024-02534-9

4/4

Fully nonlinear neuromorphic computing with linear wave scattering - Nature Physics

As the energy consumption of neural networks continues to grow, different approaches to deep learning are needed. A neuromorphic method offering nonlinear computation based on linear wave scattering can be implemented using integrated photonics.

Nature