Neural Networks

Optics and coherent light-matter systems have the potential to implement artificial neural networks (ANNs), a technology increasingly recognized for its inherent parallelism and low energy consumption, making it a suitable candidate for industrial and fundamental applications. While ANNs are predominantly electronic-based, the shift towards photonic neural networks is gaining momentum. In photonic ANNs, mathematical operations are mapped onto optical propagation characteristics, manipulated by programmable linear optics and nonlinearity. Synaptic weights in these networks are scalar and represent the pairwise connections between neurons, with the layout of interconnections mirroring matrix-vector operations where neuron inputs are the dot products of outputs from connected neurons with assigned weights.

Photonic implementations of ANNs can be categorized into three distinct types. First, free-space systems utilize diffraction and Fourier transforms, offering high scalability and the capacity to process numerous neurons simultaneously, although they face limitations in connectivity. Examples include scalable, two-layer NNs for classifying statistical Ising model phases. Second, spatial light modulators (SLMs) program linear operations, and Fourier lenses aggregate light power-encoded signals for summation. However, realizing nonlinear optical activation functions in free-space optics can be complex, as seen in systems using laser-cooled atoms with electromagnetically induced transparency. Lastly, on-chip designs employing wavelength multiplexing or beamsplitter meshes facilitate programmable all-to-all coupling, yet they require improvements in scalability. Notable advancements include an on-chip design for handwriting recognition, leveraging phase and magnitude information encoding in optical interference and using Mach–Zehnder Interferometers for various functions, including light signal modulation and complex-valued weight matrix implementation. In our research, we study the evolving landscape of photonic ANNs, highlighting their unique advantages and challenges in this promising field.