Fig. 3: Spectral recognition experiment using the PEIC.
From: Photonic edge intelligence chip for multi-modal sensing, inference and learning
a Experimental setup: An amplified spontaneous emission (ASE) source is shaped by a waveshaper into four tablet types (A, B, C, D) derived from a public dataset. The shaped spectra are amplified by an erbium-doped fiber amplifier (EDFA) and then coupled into the PEIC for classification. b (i) Example input spectra for the four tablet types. (ii) Corresponding spectral sampling results after the AWG, where each channel captures a portion of the spectrum for parallel processing. c Neural network architecture and in situ learning methods on the PEIC. (i) The neural network diagram of the PEIC. (ii) In situ supervised learning uses labeled data to update on-chip tunable devices via gradient descent. (iii) Fine-tuning with unlabeled data aligns the hardware output distribution with a reference model through a local loss function, mitigating process variations and device nonuniformities without requiring labels. d Training curves over 200 epochs and the corresponding confusion matrix. e Comparison of confusion matrices without (left) and with (right) the hardware fine-tuning procedure. After fine-tuning, the accuracy increases from 75% to 97.5%.
