Analysis of Noise-Tolerant Free Space Optical Convolution Neural Networks

Optical Neural Networks have been proposed as a promising alternative to overcome the limitations of electronic systems [1], including significant power consumption, and computational complexity, which challenge the scalability of, among other neural networks models, also Convolutional Neural Networks (CNNs) [2]. Optical Convolutional Neural Networks (OCNNs) can address these issues by leveraging optical hardware to perform convolutions on-the-fly [3]. Kernel training is usually performed digitally, neglecting optical noise, which can degrade the accuracy of OCNNs. This is especially critical for Deep OCNNs, as they are highly sensitive to both systematic noise caused by predictable or repeatable noises (lens aberration, diffraction effects, misalignment of optical components, etc.) and random noises deriving from unpredictable variations (detector noise, laser intensity fluctuations, etc,), especially when Deep OCNNs (DOCNNs) are targeted. We have recently implemented a DOCNN in the loop that achieves a 10-fold reduction in latency and a power efficiency improvement of 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> times compared to traditional DCNNs using state-of-the-art electronics [4]. This paper examines the impact of noise contributions within OCNNs, an area that has not been previously explored. In this work we explore the performance of OCNNs, implemented using non-ideal (noise effected) 4F amplitude-only OCNN for dataset classification. The correlator generates feature maps by performing dot products between kernels and input images in the frequency domain. To extend the optical correlator from a single layer to multiple convolutional layers in a DOCNN, the output images (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$out_{t}$</tex>) are stored and reused as inputs for the subsequent layers, such that <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$i n_{t+1}=o u t_t$</tex> (see Fig. 1a). This layered architecture allows the system to handle more complex tasks by enabling deeper feature extraction through successive optical convolutions [5], despite of the induced noise.