Hybrid deep learning with optical 4f correlator: simulation and performance analysis

Deep Convolutional Neural Networks (DCNNs) are renowned in image classification but are computationally demanding. This paper presents a 4f optical correlator simulation to implement a deep optical CNN (DOCNN) in the Fourier domain. The system models light propagation through optical and electro-optical components, performing not only convolution but also non-linearity and pooling in the optical domain, while the final classification is executed digitally. Results show that 7% improvement in accuracy is obtained when executing more layers in the optical domain and in a Deep OCNN fashion. This suggests that computation in the optical domain and in free space is promising when deep neural models are implemented, which is further underscored by notable improvements in latency and power consumption.