On 18 November at 12:00 p.m., the DEEC TALK “All-Optical Neural Networks: Towards Artificial Life” will take place. The event will feature Marco Piccardo, Assistant Professor of Physics at Técnico and researcher at INESC Microsistemas e Nanotecnologias (INESC MN).

Abstract

Light is good at the linear part of a neural network: passing through lenses and masks, it performs a large matrix multiplication in one pass, almost for free. The nonlinear part is usually left to electronics.

We built what may be the simplest all-optical nonlinear neural network. Two laser pulses cross in a thin silicon wafer and write a grating that exists only when both are present; two million trainable phase pixels then route the light. As an AND gate, its output is about 1’000 times brighter when both pulses arrive than when only one does, and 400 times the limit of any linear optical system. Physical neural networks are messier than artificial ones: the material sets the nonlinearity and the hardware can drift. Taming ours took a digital twin, careful calibration and full automation, which opened the door to physical AI: a language-model agent that operates the setup, reads the data and proposes changes. Working alongside us, it helped raise our benchmark over 100 times.

Computing is not the end goal. We envision feeding the neural network back into itself, so that it becomes a cavity where light updates its own state on every round trip, as in a cellular automaton. Our physical models of such cavities, with diffraction, saturable gain and noise, host patterns that move, grow from a seed, withstand perturbations and fade when the feedback is cut. We are now testing them against the four pillars that define artificial life. This is the question we are pursuing: what would it take for light to behave like life?

About the speaker

Marco is an Assistant Professor of Physics at Técnico Lisboa and a Visiting Professor at Harvard. Trained at École Polytechnique and Harvard, he works on meta-optics, structured light and ultrafast lasers, with around 60 papers, including in Nature and Science. He holds ERC and EIC grants, 7 patents in photonics, and co-founded Altitude, a photonics startup for aerospace.