2021/01/11 by Bodo Rueckauer, Connor Bybee, Rueckauer, Bodo +9 · 2 citations
Engineering · Neuroscience · #Advanced Memory and Neural Computing #Emerging Technologies (cs.ET) #FOS: Computer and information sciences #Ferroelectric and Negative Capacitance Devices #Neural dynamics and brain function
paper · pdf · doi:10.48550/arxiv.2101.04261
openalex publication_date 2021/01/11 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Spiking Neural Networks (SNNs) are a promising paradigm for efficient\nevent-driven processing of spatio-temporally sparse data streams. SNNs have\ninspired the design and can take advantage of the emerging class of\nneuromorphic processors like Intel Loihi. These novel hardware architectures\nexpose a variety of constraints that affect firmware, compiler and algorithm\ndevelopment alike. To enable rapid and flexible development of SNN algorithms\non Loihi, we developed NxTF: a programming interface derived from Keras and\ncompiler optimized for mapping deep convolutional SNNs to the multi-core Intel\nLoihi architecture. We evaluate NxTF on DNNs trained directly on spikes as well\nas models converted from traditional DNNs, processing both sparse event-based\nand dense frame-based data sets. Further, we assess the effectiveness of the\ncompiler to distribute models across a large number of cores and to compress\nmodels by exploiting Loihi's weight sharing features. Finally, we evaluate\nmodel accuracy, energy and time to solution compared to other architectures.\nThe compiler achieves near optimal resource utilization of 80% across 16 Loihi\nchips for a 28-layer, 4M parameter MobileNet model with input size 128x128. In\naddition, we report the lowest error rate of 8.52% for the CIFAR-10 dataset on\nneuromorphic hardware, using an off-the-shelf MobileNet.\n