Neuromorphic processors commonly execute spiking neural networks (SNN) models to obtain high energy efficiency. Compared to standard SNNs, liquid state machine (LSM), the spiking variant of reservoir computing, exhibits advantages in image classification, and are especially more promising in speech recognition. However, LSM-based neuromorphic processors suffer from weight storage overhead in resource-constrained edge applications. To address this, we propose ghost reservoir: a memory-efficient LSM based neuromorphic processor enabling on-chip spike-time-dependent plasticity-based backpropagation (BP-STDP) learning. For the LSM reservoir layer, we adopt a stateless-neuron model and an on-the-fly weight re-generation scheme to avoid the storages for both membrane potentials and weights. For the readout learning layer, a stochastic weight update approach is implemented to reduce the memory bit-width. These techniques contribute to an aggregate of a 482.5 KB on-chip memory reduction on the FPGA prototype. Implemented on the very-low-cost Xilinx Zynq-7010 device, our prototype achieved real-time processing, demonstrated comparably high on-chip learning accuracies of 94.93%, 84.65%, and 92% on the MNIST, N-MNIST, and FSDD datasets, respectively. These experimental results indicate that our LSM-based lightweight neuromorphic design is quite suitable for speech and visual recognition tasks in many resource-constrained edge intelligent applications.
11 June 2024 in IEEE Transactions on Biomedical Circuits and Systems
This paper presents a digital edge neuromorphic spiking neural network (SNN) processor chip for a variety of edge intelligent cognitive applications. This processor allows high-speed, high-accuracy and fully on-chip spike-timing-based multi-layer SNN learning. It is characteristic of hierarchical multi-core architecture, event-driven processing paradigm, meta-crossbar for efficient spike communication, and hybrid and reconfigurable parallelism. A prototype chip occupying an active silicon area of 7.2 mm 2 was fabricated using a 65-nm 1P9M CMOS process. when running a 256-256-256-256-200 4-layer fully-connected SNN on downscaled 16 × 16 MNIST images. it typically achieved a high-speed throughput of 802 and 2270 frames/s for on-chip learning and inference, respectively, with a relatively low power dissipation of around 61 mW at a 100 MHz clock rate under a 1.0V core power supply, Our on-chip learning results in comparably high visual recognition accuracies of 96.06%, 83.38%, 84.53%, 99.22% and 100% on the MNIST, Fashion-MNIST, ETH-80, Yale-10 and ORL-10 datasets, respectively. In addition, we have successfully applied our neuromorphic chip to demonstrate high-resolution satellite cloud image segmentation and non-visual tasks including olfactory classification and textural news categorization. These results indicate that our neuromorphic chip is suitable for various intelligent edge systems under restricted cost, energy and latency budgets while requiring in-situ self-adaptative learning capability.
27 February 2015 in Research on Chemical Intermediates
This brief proposes an edge neuromorphic hardware design for real-time energy-efficient applications. It is capable of fast on-chip learning on compressive sensed spikes utilizing an error-triggered learning mechanism. Our hardware architecture consists of two event-driven arrays of parallel computing cores, and adopts fine-grained pipeline circuits to boost processing speed. The architecture has good scalability and provides a flexible trade-off among processing speed, recognition accuracy and resource cost. Our design was prototyped on the very-low-cost Xilinx Zynq-7010 FPGA chip on a Zybo platform. The FPGA prototype realized a high processing speed of 4843 and 4931 frames/s with a high energy efficiency of $31.39~\mu \text{J}$ /image and $30.83~\mu \text{J}$ /image on the MNIST image set during training and inference, respectively. The prototype reached comparably high on-chip learning accuracies across a serial of datasets including both static images as well as spike event streams, e.g., 95.02% on the MNIST images and 84.46% on the N-MNIST streams.
30 October 2024 in IEEE Transactions on Circuits and Systems II: Express Briefs
Neuromorphic processors are promising candidates for energy-constrained intelligent systems, as they emulate cortical computations via spatiotemporally sparse binary spikes. However, achieving high-accuracy, high-throughput and cost-efficient neuromorphic processing remains challenging. To fully utilize hardware resources for performance improvement, we propose a multi-core neuromorphic architecture characteristic of a uniform neuron-core mapping scheme and a layer-wise event-batch-based parallel processing paradigm. These techniques ensure highly balanced cross-core workloads regardless of actual mapped neural network topologies as well as unpredictable input and internally generated spike counts varying from sample to sample. An FPGA prototype of our neuromorphic processor was implemented. It exhibited comparably high on-chip learning accuracies on various visual and non-visual benchmarks, high learning/inference frame rates (low processing latencies), with a moderate amount of logic and memory resource consumptions.
30 October 2024 in Computers and Electrical Engineering