Noise-mitigation strategies in physical feedforward neural networks

Area of Science:

• Artificial Intelligence Hardware
• Neuromorphic Engineering
• Signal Processing

Background:

• Physical neural networks (PNNs) are a promising hardware for next-generation AI, but are susceptible to noise inherent in analog systems.
• Unlike digital systems, PNNs lack a high signal-to-noise ratio, necessitating efficient noise mitigation strategies.
• Hardware-efficient noise reduction is crucial for realizing the potential of PNNs.

Purpose of the Study:

• To introduce and analyze novel noise-mitigation approaches for physical neural networks.
• To develop synergistic strategies for suppressing both uncorrelated and correlated noise in analog hardware.
• To demonstrate the effectiveness of these strategies in improving PNN performance.

Main Methods:

• Analytical derivations to understand noise suppression mechanisms.
• Introduction of 'ghost neurons' to mitigate correlated noise.
• Implementation of neuron population pooling to reduce uncorrelated noise.
• Testing the combined noise-mitigation strategy on a handwritten digit classification task.

Main Results:

• Intra-layer connections were shown to fully suppress uncorrelated noise under specific conditions.
• The proposed ghost neuron and pooling strategies effectively mitigated correlated and uncorrelated noise, respectively.
• A fourfold improvement in output signal-to-noise ratio was achieved.
• Classification accuracy improved from 92.07% (noisy) to 97.49%, matching the noise-free network's 97.54%.

Conclusions:

• A general noise-mitigation strategy leveraging inherent statistical properties of analog hardware was developed.
• The combined approach significantly enhances the robustness and accuracy of physical neural networks.
• This work provides a pathway for developing reliable and high-performance AI hardware.