Efficient training for the hybrid optical diffractive deep neural network

As a new emerging machine learning mechanism, optical diffractive deep neural network ((ODNN)-N-2) has been intensively studied recently due to its incomparable advantages on speed and power efficiency. However, the training process of the (ODNN)-N-2 with traditional back-propagation (BP) method is always time-consuming. Here, we introduce the biologically plausible training methods without feedback to accelerate the training process of the hybrid (ODNN)-N-2. Direct feedback alignment (DFA), error-sign-based DFA (sDFA) and direct random target projection (DRTP) are utilized and evaluated in the training process of the hybrid (ODNN)-N-2 respectively. For the hybrid (ODNN)-N-2 with 20 diffractive layers, about 160x (DFA; CPU), 30x (DFA; GPU), 170x (sDFA; CPU), 32x (sDFA; GPU), 158x (DRTP; CPU) and 32x (DRTP; GPU) accelerations are achieved respectively without significant loss of accuracy, compared with the training process using BP method on CPU or GPU.