The sigmoid function is a widely used nonlinear activation function in neural networks. In this article, we present a modular approximation methodology for efficient fixed-point hardware implementation of the sigmoid function. Our design consists of three modules: piecewise linear (PWL) approximation as the initial solution, Taylor series approximation of the exponential function, and Newton–Raphson method-based approximation as the final solution. Its modularity enables the designer to flexibly choose the most appropriate approximation method for each module separately. Performance evaluation results indicate that our work strikes an appropriate balance among the objectives of approximation accuracy, hardware resource utilization, and performance.

It is challenging to train deep spiking neural networks (SNNs) directly due to the difficulties associated with the nondifferentiable neuron model. In this work, an end-to-end learning algorithm based on discrete current-based leaky integrate-and-fire (C-LIF) neuron model and surrogate gradient is proposed to leverage the encoder–decoder network architecture to train deep SNNs directly. The proposed algorithm is capable of learning deep spatiotemporal features relying on current time step only, and several acceleration techniques including backward phase skipping and layerwise FreezeOut are proposed to accelerate the training. Experimental results show that the proposed learning algorithm achieved the classification accuracies of 98.40% and 95.83% on the dynamic neuromorphic data sets MNIST-DVS and DVS-Gestures, respectively, and of 99.58% and 95.97% on static vision data sets MNIST and SVHN, respectively, which are comparable to the existing state-of-the-art results. The training speed was accelerated by up to 49.5% on MNIST and 36.6% on DVS-Gestures with the proposed acceleration techniques while maintaining the same level of accuracy.

Thanks to their event-driven nature, Spiking Neural Networks (SNNs) enjoy great energy-efficiency and are believed to be a viable solution to the power wall problem. Due to difficulty in direct training, researchers have proposed indirect conversion-based methods, which allow SNNs to achieve comparable accuracy to their original non-spiking counterpart. However, most of these methods focus on classification problems; object detection with SNN, which involves a more challenging regression problem, remains an open research problem. The only existing SNN-based object detection method Spiking YOLO relied on a modified Integrate-and-Fire (IF) neuron model with two firing thresholds, which is not supported by most neuromorphic hardware. In order to run object detection on neuromorphic hardware, we propose a Region-based SNN (R-SNN) with widely adopted IF neurons. The bounding box regressor uses mirror output neurons along with the original output neurons to represent both positive and negative bounding box offsets, and decodes the output spike trains with a simple bounding box recovery algorithm to recover real-valued bounding box offsets. Moreover, we deploy our R-SNN on our neuromorphic computing system “Darwin Mouse” to develop a low-power object detection application, demonstrating the feasibility of applying our method to realworld neuromorphic hardware with limited arithmetic precision. Experiments show that our R-SNN achieves a mean average precision (mAP) of 63.1% on VOC 2007, improving the state of the art (achieved by Spiking YOLO) by more than 11%.

Spiking Neural Networks (SNNs) have great advantages in discrete event data processing because of their binary digital computation form. However, due to the limitation of the current structures of SNNs, the original event data needs to be preprocessed to reduce the time calculation steps and information redundancy. The traditional methods of dividing data into frames lead to the loss of a large amount of time information. In this paper, we proposed an efficient Recurrent Spiking Neural Network (RSNN) to reduce the time domain information loss of original slice samples with the spiking based neural dynamics for processing the dynamic spatial-temporal information. By constructing the Recurrent Spiking Neural Network model, the recurrent structure was used to preprocess slices before it was further input into the spiking structure to enhance the time correlation between slices. In addition, in order to match the two-dimensional spatial structure of data sample frames efficiently, this paper adapts a variation of structures of the recurrent neural network, named Convolution LSTM (CONLSTM). Through experiments on event based datasets such as DVS128-Gesture and CIFAR10-DVS, we find that the proposed model could not only behave better than some other spiking based models but also save energy and power consumption which paves the way for practical applications of neuromorphic hardware.

The traffic patterns of spiking neural networks (SNNs) exhibit high variability and stochastic, leading to the emergence of elevated traffic hotspots on the network-on-chip (NoC)-based neuromorphic processors. Predicting the occurrence of hotspots remains one of the most challenging issues in NoC design. This article presents the first attempt toward traffic hotspot prediction by utilizing liquid state machine (HP-LSM). The predictor extracts essential information reflecting the current state of the NoC to predict potential routing hotspots in the subsequent time step. Furthermore, we designed the hardware architecture for HP-LSM, which incorporates leaky-integrate-and-fire (LIF) neurons with configurable biological parameters. Meanwhile, we introduce a novel hotspot-aware path-based multicast (HaPM) routing algorithm that utilizes advanced knowledge acquired from HP-LSM to guide packet routing throughout the network, aiming to improve the performance of NoC. Results indicate that the HP-LSM can forecast hotspot formation with an accuracy up to 89.36% and 90.19% for two spiking-based datasets, respectively. The hardware experiment results demonstrate a 92.03% reduction in the average execution time of zero skipping compared with nonzero skipping. Moreover, the HP-LSM exhibits a reduction of up to 79.30% in the number of neurons compared with other related SNN predictor models. The experiments reveal a reduction of 73.67% and 53.42% in the average length of the multicast path when compared with dual-path (DP) or multipath (MP) multicast routing. The HaPM demonstrates improved performance in terms of average latency and throughput compared with DP, MP, and path-based multicast (PbM) multicast routing.