Bringing uncertainty quantification to the extreme-edge with memristor-based Bayesian neural networks.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Device Physics

Background:

• Safety-critical applications require accurate decisions from limited, noisy data.
• Bayesian neural networks (BNNs) provide predictive uncertainty but are computationally intensive.
• Memristors offer a probabilistic hardware platform for BNN implementation.

Purpose of the Study:

• To develop a memristor-based Bayesian neural network overcoming device physics limitations.
• To achieve energy-efficient in-memory computing for BNNs.
• To demonstrate the practical application of this technology in medical diagnosis.

Main Methods:

• Implemented a variational inference training method augmented with a "technological loss" function.
• Integrated memristor physics into the BNN training process.
• Programmed a BNN on 75 memristor crossbar arrays with CMOS periphery for in-memory computing.

Main Results:

• The memristor-based BNN accurately classified heartbeats and estimated prediction certainty.
• Achieved orders-of-magnitude improvement in inference energy efficiency compared to conventional hardware.
• Demonstrated successful in-memory computing for a probabilistic neural network.

Conclusions:

• Memristor-based BNNs offer a highly energy-efficient solution for safety-critical AI.
• The "technological loss" effectively bridges memristor device physics with BNN requirements.
• This approach paves the way for advanced, low-power neuromorphic computing systems.