AE-D2NN: Autoencoder in Diffractive Neural Network Form
Peijie Feng, Zongkun Zhang, Mingzhe Chong, Yunhua Tan · 2024
The development and potency of deep neural networks have revolutionized numerous fields, demonstrating unprecedented capabilities in complex tasks. Despite this, the advancement of deep neural networks is increasingly constrained by the computational limits of hardware. In response, many researchers are shifting their focus to optical neural networks for higher computation efficiency and speed, seeking to harness the ultra-fast processing capabilities of light and its potential for parallel computing. Among them, the diffractive deep neural network stands out, uniquely utilizing the principles of light diffraction and interference for data processing. However, its application has predominantly focused on functioning as a multi-layer linear perceptron for classification tasks utilizing light waves, leaving its potential in other deep learning tasks and models largely untapped. Autoencoder, as a fundamental part of generative models in deep learning, excel in unsupervised learning for efficient data encoding, showcasing their potential in data compression and reconstruction. Therefore, in this article, we explored the combination of autoencoder and diffractive deep neural network, and proposed the AE-D2NN model. We adopted a knowledge distillation strategy, using a teacher AE model to train the student AE-D2NN model on the MNIST and Fashion MNIST datasets. The trained model is capable of performing image compression and reconstruction, which can also be interpreted as beam focusing and holographic imaging functions. The AE-D2NN model is a twin of the autoencoder model under the framework of diffractive neural networks, representing an initial exploration of the integration of diffractive neural network with generative models. This work may also inspire new approaches to electromagnetic wave manipulation.