Analog Computation with Oscillatory Neural Networks

Inspired by the brain, neuromorphic computing aims to circumvent the von Neumann’s bottleneck by bringing memory together with processing units. Besides avoiding undesirable data transfers, such physics-based approach can further reduce the computational cost for low-power IoT applications by encoding variables into analog physical quantities. Oscillatory neural networks (ONNs) are promising neuromorphic systems where the computational process emerges from the natural synchronization of coupled oscillators (neurons) in continuous-time and without any underlying algorithm. However, due to their analog nature, ONNs are challenging to implement at a large-scale. In this talk, I will present a novel mixed-signal ONN architecture that computes in the analog domain while propagating information digitally to facilitate scaling and to enable I/O communications. We highlight the potential of the proposed architecture with a prototype TSMC 65-nm chip containing 16 relaxation oscillators (neurons) and 256 programmable synapses with 5 bits of resolution. The IC core and IOs consume 167µW and 6.2 mW, respectively. When programmed to solve NP-hard combinatorial optimization problems (COPs) such as Max-cut, the ONN finds larger cuts than 0.95x the state-of-the-art Goemans-Williamson algorithm (GW) executed on a CPU for a 1.5 104 x runtime improvement. Owing to its digital signal propagation, the proposed ONN can be scaled up and provide a low-power alternative to CPUs when solving COPs in IoT devices.