Development of online learning technique for magneto-optical diffractive deep neural networks

Abstract We demonstrate an online learning of a single-layer magneto-optical diffractive deep neural network (MO-D 2 NN) for the classification of handwritten digits. MO-D 2 NN realizes light modulation by utilizing Faraday effect of magnetic materials, allowing operation in the visible light and reconfigurability using rewritable magnetic domains that play a role of binary weights. An online learning experiment was demonstrated for a single image as input. A Y 0.5 Bi 2.5 Fe 4 GaO 12 thin film with a Faraday rotation of 3.3° at a wavelength of 532 nm was used as the hidden layer. A derivative-free optimization algorithm based on the Monte Carlo method was implemented for the online learning, adjusting the magnetic domains iteratively to minimize cross-entropy loss. Following the developed Monte Carlo algorithm, the magnetic domain pattern was updated using thermomagnetic recording technique, and the cross-entropy loss was sufficiently reduced, resulting that the input image was classified appropriately.