Investigating Dimensionality Reduction Strategies for Future Optical Neural Computing Systems

Nishtha Gandhi

Abstract: Artificial intelligence workloads continue to grow in computational complexity, motivating research into alternative computing paradigms capable of delivering improved scalability and energy efficiency. Optical neural computing has emerged as a promising candidate due to the inherent parallelism of light and the possibility of performing large-scale linear operations at extremely high speed. However, practical realization of optical neural systems remains challenging because high-dimensional data representations often require a correspondingly large number of physical optical modes. This study investigates whether intelligent feature-compression techniques can reduce dimensionality while preserving classification performance. Using handwritten digit recognition as a benchmark task, four dimensionality reduction approaches are evaluated: direct resolution reduction, learned bottleneck compression, Principal Component Analysis (PCA), and autoencoder-based encoding. The work establishes quantitative dimensionality targets for future optical computing systems and provides the theoretical foundation for future Jacobi Time-Wave Packet implementations.

Keywords: Optical Neural Computing, Machine Learning, Neural Networks, Feature Compression, Dimensionality Reduction, Principal Component Analysis (PCA), Autoencoders, Representation Learning, MNIST, Handwritten Digit Recognition, Optical Machine Learning, Jacobi Time-Wave Packets