Abstract
With recent rapid advances in photonic integrated circuits, it has been demonstrated that programmable photonic chips can be used to implement artificial neural networks. Convolutional neural networks (CNN) are a class of deep learning methods that have been highly successful in applications such as image classification and speech processing. We present an architecture to implement a photonic CNN using the Fourier transform property of integrated star couplers. We show, in computer simulation, high accuracy image classification using the MNIST dataset. We also model component imperfections in photonic CNN and show that the performance degradation can be recovered in a programmable chip. Our proposed architecture provides a large reduction in physical footprint compared to current implementations as it utilizes the natural advantages of optics and hence offers a scalable pathway towards integrated photonic deep learning processors.
Highlights
Deep learning methods like Convolutional neural networks (CNN) have received a huge amount of interest from the research community as well as the general public after it was shown to approach human level performance in image recognition tasks [1]
Since the bulk of the trainable parameters in the photonic convolutional neural network (PCNN) come from the fully-connected layers, whereas the parameters in the convolution layers only grow linearly with the size of the input, we studied the trade-off between performance and complexity by being more aggressive with the pooling layers
To study the effects of such imperfections, we evaluated the degradation of the MNIST classification accuracy of a pre-trained PCNN-784 by introducing both be found in the literature [13], [55], [59], [60]
Summary
Deep learning methods like CNNs have received a huge amount of interest from the research community as well as the general public after it was shown to approach human level performance in image recognition tasks [1] This breakthrough was due in part to the availibility of fast graphical processing units (GPUs) that greatly accelerated the implementation of deep neural networks [2]. Diffractive optics systems have demonstrated a high degree of success in image classification tasks [16], [17] Such systems are constructed using sequential amplitude and phase masks, with each individual pixel in the masks being a trainable parameter in a deep learning optimization algorithm. We present the performance of the photonic CNN on various datasets, study the performance degradation with imperfections and provide discussion on scalability and possible future directions
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
