December 2022
Spotlight Summary by Daniel Brunner
Optical processor for a binarized neural network
Neural networks (NNs) enabled computing tasks that previously appeared reserved to abstract human intelligence. These computing concepts continue to surprise, for example by reaching almost human-quality language translations. Yet, hardware remains ill-suited for implementing NNs. An astronomical energy consumption, chiefly caused by inefficient implementations of a NN’s weighted connections, threatens to limit progress. In this context, Huang and Yao report an approach to photonically implement this operation. Superior performance in parallel communication makes photonic NNs promising alternatives. However, programmability of nonlinear circuits, which is essential for NNs, remains difficult to realize optically. In “Optical processor for a binarized neural network”, the authors implement programmable binary weights (-1, +1) and a neuron’s nonlinearity using a single commercial component, a dual-drive Mach-Zehnder modulator. A multi-layer NN is hardware implemented, reaching surprisingly high performance metrics considering the restriction to binary-only weights. In order to exploit the potential of photonics, the concept should be expanded in order to remove the serialized nature that it fundamentally relies upon in the current version.
You must log in to add comments.
Add Comment
You must log in to add comments.
Article Information
Optical processor for a binarized neural network
Long Huang and Jianping Yao
Opt. Lett. 47(15) 3892-3895 (2022) View: Abstract | HTML | PDF