Programmable Photonic Extreme Learning Machines

TL;DR

This work tackles the training bottleneck of photonic neural networks by employing a programmable photonic extreme learning machine (PPELM) that uses a hexagonal waveguide mesh to realize a random feature map on-chip, with nonlinear activation implemented by integrated photodetectors and the final regression performed digitally. The authors demonstrate on-chip programmable input encoding, a random projection layer, and a trainable output layer across three classification tasks, revealing how increasing hidden nodes reduces variance but raises hardware demands. To boost performance without expanding hardware, two strategies are shown: differential evolution to optimize the random hidden layer and a WDM-based ensemble that trains multiple models in parallel across wavelengths and ensembles their predictions. The results indicate that programmable photonic processors can train competitive ML models on compact hardware, offering a viable path toward scalable, low-latency photonic accelerators with tunable accuracy enhancements.

Abstract

Photonic neural networks offer a promising alternative to traditional electronic systems for machine learning accelerators due to their low latency and energy efficiency. However, the challenge of implementing the backpropagation algorithm during training has limited their development. To address this, alternative machine learning schemes, such as extreme learning machines (ELMs), have been proposed. ELMs use a random hidden layer to increase the feature space dimensionality, requiring only the output layer to be trained through linear regression, thus reducing training complexity. Here, we experimentally demonstrate a programmable photonic extreme learning machine (PPELM) using a hexagonal waveguide mesh, and which enables to program directly on chip the input feature vector and the random hidden layer. Our system also permits to apply the nonlinearity directly on-chip by using the systems integrated photodetecting elements. Using the PPELM we solved successfully three different complex classification tasks. Additioanlly, we also propose and demonstrate two techniques to increase the accuracy of the models and reduce their variability using an evolutionary algorithm and a wavelength division multiplexing approach, obtaining excellent performance. Our results show that programmable photonic processors may become a feasible way to train competitive machine learning models on a versatile and compact platform.

Figures (6)

• Figure 1: Programmable photonic extreme learning machines: a Diagram of the architecture of the extreme learning machines, b Schematic of the experimental setup for the implementation of programmable photonic extreme learning machines. The hexagonal mesh is programmed to multiply the input data by the random matrix and apply the nonlinear function using the square law of the photodetectors. The trainable weight matrix is multiplied on the CPU and, c different states of the PUCs on the hexagonal waveguide mesh.
• Figure 2: Accuracy of the of the PPELM in the a Header recognition task, the b Iris Flower and c Banknote Authentication Datasets Classification tasks when using 4, 6, 8 and 10 hidden nodes. In blue are the results of the train and in orange of the test set for 40 different random initializations.
• Figure 3: Evolution of the training and validation accuracy during 35 iterations of the DE-PPELM when using a 4, b 6, c 8, d 10 hidden nodes in the Iris Flower Dataset classification task.
• Figure 4: Evolution of the training and validation accuracy during 35 iterations of the DE-PPELM when using a 4, b 6, c 8, d 10 hidden nodes in the Banknote authentication dataset classification task.
• Figure 5: Ensembled extreme learning machines using WDM.