Forward-forward training of an optical neural network

Area of Science:

• Optics
• Artificial Intelligence
• Machine Learning

Background:

• Neural networks (NNs) require significant computational resources, driving demand for faster, energy-efficient hardware.
• Optical platforms offer potential for efficient NN hardware but face training challenges due to difficulties in characterizing physical systems for backpropagation.
• The forward-forward algorithm (FFA) presents a novel training approach that bypasses the need for differentiable physical system characterization.

Purpose of the Study:

• To demonstrate the feasibility of the forward-forward algorithm (FFA) for training optical neural networks (NNs) using a physical optical system.
• To explore the potential of leveraging physical transformations within NN architectures for performance enhancement.
• To address the challenges of training programmable optical NN layers without relying on traditional error backpropagation.

Main Methods:

• An experimental setup utilizing multimode nonlinear wave propagation in an optical fiber was employed.
• The forward-forward algorithm (FFA) was applied to train a multilayer NN architecture incorporating optical transforms.
• The FFA's local loss function enabled weight updates without backpropagating error signals, suitable for analog hardware.

Main Results:

• The feasibility of the FFA for training an optical NN system was successfully demonstrated.
• Incorporating optical transforms within NN architectures trained by FFA led to performance improvements.
• Effective training was achieved even with a limited number of trainable weights, highlighting the algorithm's efficiency.

Conclusions:

• The forward-forward algorithm (FFA) offers a viable and efficient method for training optical neural networks (NNs).
• This approach overcomes key limitations associated with training physical NN hardware, particularly the need for differentiable models.
• The study opens new avenues for developing low-power, high-performance optical NN hardware by integrating physical transformations and novel training algorithms.