In-Sensor Learning on Magnetic Tunnel Junction Devices Using the Forward-Forward Algorithm

Aiming at the real-time processing requirements in edge computing devices, we construct in-sensor computing units based on magnetic tunnel junction (MTJ) arrays, which realizes in-situ computation of vector-matrix multiplication with parameters dynamically programmed to adapt to different tasks. Different from the previously used back-propagation (BP) algorithm, we innovatively introduce the forward-forward (FF) algorithm, which utilizes the positive and negative samples “Goodness” optimization mechanism to reduce the computational complexity. In the noisy image classification task, the FF algorithm reduces the training time by more than 90% compared to the BP algorithm and significantly improves the convergence speed when perfect classification results are achieved.