Validation of Photonic Neural Networks in Health Scenarios

Photonic hardware represents a promising alternative to speed-up Neural Network (NN) computations, outperforming electronic counterparts in terms of speed, energy consumption and computing density. In this paper we exploit a Photonic-Aware Neural Network (PANN) architecture with unipolar and bipolar weight implementations, considering ReLU and photonic sigmoid as candidate activation functions to solve a heartbeat sound classification task. Results indicate that increasing the bitwidths during quantization improves the F1-score. The use of bipolar implementation for weight choice demonstrates better performance. ReLU is identified as a better nonlinearity. Finally, a multi-resolution scenario in the bipolar photonic-sigmoid experiment is evaluated, revealing that incorporating multi-resolution does not enhance the model's generalization ability if the bitwidth for the first layer remains fixed. However, the importance of the highest bitwidth at the NN inputs is highlighted.