AxNN: Energy-efficient neuromorphic systems using approximate computing

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