Sample-efficient, low-light image sensing through Eigentask Learning: Part 1 (Theory)

The computational capacity of physical neural networks (PNNs) is inherently constrained by ubiquitous readout noise, accentuating the need for optimal information extraction especially under low-power operation. In this joint theory-experiment study, we demonstrate the first practical application to optical PNNs of a sample-efficient supervised machine learning technique known as Eigentask Learning (EL) [1] - to high-dimensional classical optical sensor systems operating at low photon numbers. In the first part of this talk, we introduce EL alongside other noise reduction methods, including principal component analysis and low pass filtering, which are commonly used to process data collected from optical devices operating under low signal powers. Being a similarly linear framework, EL is shown to maintain the simplicity of these standard approaches, but remarkably provides a robust and consistent improvement in performance across various metrics, including a substantial reduction in features needed for high-accuracy training, faster convergence in optimization, and improved maximum accuracy in learning. This improvement holds across both simple and complex photonic devices, as we show via simulations that form a precursor to experimental demonstrations in the second part of this talk.

[1] Hu et al. Phys. Rev. X 13, 041020 (2023).