AiPT Research Seminars

The AiPT research seminars bring together international experts and AiPT staff, fostering an environment of knowledge sharing and discovery across a range of diverse topics. The AiPT research seminars continue to promote scientific dialogue and engagement among the global research community, ensuring that AiPT remains at the forefront of innovation and discovery in the fields of photonics and telecommunications.

AIPT Upcoming Seminars in 2026

Speaker: Andrei V. Ermolaev
Affiliation: FEMTO-ST Institute (CNRS)
Title: Power law scaling for classification performance in physical neural networks
Time
: 15 October, Thursday from 11:00am to 12:00pm
Venue: NW708, MS Teams

Abstract:Physical neural networks (PNNs) utilise the intrinsic complexity of physical systems to perform computations, potentially achieving speed and energy efficiency inaccessible to conventional digital hardware. However, to date, no general methodology has been developed to quantitatively assess and predict the performance of such computing across diverse substrates. In this talk, I will introduce the Hotelling Trace Criterion (HTC), a task-dependent measure of the separability of PNN states that can be estimated without training. We have recently demonstrated that HTC can accurately predict the classification performance of PNNs based on highly nonlinear optical fibres, vertical-cavity surface-emitting lasers, as well as networks of coupled nonlinear oscillators (CNON), for benchmark tasks of varying complexity. Classification loss has been shown to follow a power-law relationship with HTC, with Pearson correlation coefficients of |r| ≈ 0.99 for MNIST and ≈ 0.97 for Fashion MNIST. It is noteworthy that experimental and simulated data from physically distinct systems collapse onto a single scaling curve determined by the task rather than by substrate. Applying HTC layer by layer during training of the multilayer CNON PNN further reveals that gradient-based optimisation distributes representational capacity unevenly across PNN layers, highlighting that HTC provides a quantitative diagnostic of training efficiency that is invisible to standard loss monitoring. These results establish HTC as a substrate-agnostic metric for comparing and scaling PNNs, advancing the field towards a complete theory connecting fundamental hardware parameters to task performance through universal scaling laws.

Past Seminars

AiPT Seminar Series

The AiPT team organises a series of scientific seminars covering a wide range of topics, from experimental and theoretical challenges in photonics to industrial applications. If you are interested in collaborating with our team and would like to deliver a talk, please contact the AiPT Seminar Chair, Dr Auro Perego, at a.perego1@aston.ac.uk