In this paper, we present a survey of recent works in developing neuromorphic or neuro-inspired hardware systems.In particular, we focus on those systems which can either learn from data in an unsupervised or online supervised manner. We present algorithms and architectures developed specially to support on-chip learning. Emphasis is placed on hardware friendly modifications of standard algorithms, such as backpropagation, as well as novel algorithms, such as structural plasticity, developed specially for low-resolution synapses. We cover works related to both spike-based and more traditional non-spike-based algorithms. This is followed by developments in novel devices, such as floating-gate MOS, memristors, and spintronic devices. CMOS circuit innovations for on-chip learning and CMOS interface circuits for post-CMOS devices, such as memristors, are presented. Common architectures, such as crossbar or island style arrays, are discussed, along with their relative merits and demerits. Finally, we present some possible applications of neuromorphic hardware, such as brain-machine interfaces, robotics, etc., and identify future research trends in the field.

No abstract

Some recent work revealed that deep neural networks (DNNs) are vulnerable to so-called adversarial attacks where input examples are intentionally perturbed to fool DNNs. In this work, we revisit the DNN training process that includes adversarial examples into the training dataset so as to improve DNN's resilience to adversarial attacks, namely, adversarial training. Our experiments show that different adversarial strengths, i.e., perturbation levels of adversarial examples, have different working ranges to resist the attacks. Based on the observation, we propose a multi-strength adversarial training method (MAT) that combines the adversarial training examples with different adversarial strengths to defend adversarial attacks. Two training structures-mixed MAT and parallel MAT-are developed to facilitate the tradeoffs between training time and hardware cost.Our results show that MAT can substantially minimize the accuracy degradation of deep learning systems to adversarial attacks on MNIST, CIFAR-10, CIFAR-100, and SVHN. The tradeoffs between training time, robustness, and hardware cost are also well discussed on a FPGA platform.Index Terms-neural network, adversarial example, adversarial attack, adversarial training, FPGA

No abstract

No abstract