Unfair Photonics? How Encoding May Lead to Bias in Photonic Neural Networks

Artificial Intelligence (AI) systems have reached human-level performance in numerous tasks and are now applied across a variety of fields. As these applications evolve and increase in complexity, they create a growing demand for computing solutions that offer both high speed and energy efficiency. In this context, integrated photonic circuits are emerging as a promising alternative to traditional electronic systems, particularly for AI implementations. Photonic Neural Networks (PNNs) have the potential, for instance, to operate at greater speeds and with significantly lower power consumption than their electronic counterparts. PNNs shift computations to the optical domain, using photonic hardware to process data encoded in light. A common circuit design uses meshes of interferometers, which can be configured through their phase shifters to perform matrix-vector multiplications. This enables PNNs to act as complex-valued neural networks, and
multiple real-valued features to be accommodated into a same complex-valued input. Combining features in such way reduces input dimensionality, resulting in more compact and scalable integrated chip designs. However, it can also influence the model’s performance, as the choice of encoding determines the relationships of features. Here, we explore how these fixed relationships can render PNNs more ‘rigid’ and change which features are prioritized during its decision-making process. We analyze different encoding schemes and feature pairings, illustrating how this rigidity may pose challenges in applications where unbiased outcomes are essential, as it can unintentionally amplify or reduce the importance of features.