AxNN: energy-efficient neuromorphic systems using approximate computing

摘要: Neuromorphic algorithms, which are comprised of highly complex, large-scale networks artificial neurons, increasingly used for a variety recognition, classification, search and vision tasks. However, their computational energy requirements can be quite high, hence energy-efficient implementation is great interest. We propose new approach to design hardware implementations neural (NNs) using approximate computing. Our work motivated by the observations that (i) NNs in applications where less-than-perfect results acceptable, often inevitable, (ii) they resilient inexactness many (but not all) constituent computations. make two key contributions. First, we method transform any given NN into an Approximate Neural Network (AxNN). This performed adapting backpropagation technique, commonly train these networks, quantify impact approximating each neuron overall network quality (e.g., classification accuracy), selectively those neurons least. Further, observation training naturally error-healing process mitigate approximations neurons. Therefore, incrementally retrain with in-place, reclaiming significant portion ceded approximations. As second contribution, programmable quality-configurable neuromorphic processing engine (qcNPE), utilizes arrays specialized elements execute computations dynamically configurable accuracies AxNNs from diverse applications. evaluated proposed constructing AXNNs 6 recognition (ranging complexity 12–47,818 160–3,155,968 connections) executing them on different platforms - qcNPE containing 272 45nm technology commodity Intel Xeon server. demonstrate 1.14X–1.92X benefits virtually no loss (< 0.5%) output quality, even higher improvements (upto 2.3X) when some 7.5%) acceptable.