Energy efficient training of private recommendation systems using multi-armed bandit models and analog in-memory computing

Abstract Recommendation systems, the heart of the consumer electronics industry’s profitability, train deep neural networks to deliver compelling individualized content based on complex user behavior. However, AI training today needs massive and energy-hungry data-center infrastructure, making recommendations extremely expensive. Therefore, it is crucial to find a model-algorithm combination co-designed with the constraints of low-power edge hardware. Here, using an index policy for solving the ‘restless multi-armed bandit’ decision framework, we address the challenge of learning a part of the user behavior via a privacy-preserving numerical algorithm. We co-design the model and algorithm to tolerate and exploit the nonidealities (e.g., noise and reduced precision) of low-power edge hardware. We train a recommendation model with up to 10 contents (arms) entirely in-situ on integrated 12 Mb analog-digital hybrid crossbars of resistive random-access memory. Our benchmarking shows an energy advantage of 100× relative to state-of-the-art GPU-based systems, demonstrating a clear potential for AI-driven efficient and private recommendations.