Autonomously Reconfigurable Artificial Neural Network on a Chip

Artificial neural network (ANN), an established bio-inspired computing paradigm, has proved very effective in a variety of real-world problems and particularly useful for various emerging biomedical applications using specialized ANN hardware. Unfortunately, these ANN-based systems are increasingly vulnerable to both transient and permanent faults due to unrelenting advances in CMOS technology scaling, which sometimes can be catastrophic. The considerable resource and energy consumption and the lack of dynamic adaptability make conventional fault-tolerant techniques unsuitable for future portable medical solutions. Inspired by the self-healing and self-recovery mechanisms of human nervous system, this research seeks to address reliability issues of ANN-based hardware by proposing an Autonomously Reconfigurable Artificial Neural Network (ARANN) architectural framework. Leveraging the homogeneous structural characteristics of neural networks, ARANN is capable of adapting its structures and operations, both algorithmically and microarchitecturally, to react to unexpected neuron failures. Specifically, we propose three key techniques --- Distributed ANN, Decoupled Virtual-to-Physical Neuron Mapping, and Dual-Layer Synchronization --- to achieve cost-effective structural adaptation and ensure accurate system recovery. Moreover, an ARANN-enabled self-optimizing workflow is presented to adaptively explore a "Pareto-optimal" neural network structure for a given application, on the fly. Implemented and demonstrated on a Virtex-5 FPGA, ARANN can cover and adapt 93% chip area (neurons) with less than 1% chip overhead and O(n) reconfiguration latency. A detailed performance analysis has been completed based on various recovery scenarios.