NxTF: An API and Compiler for Deep Spiking Neural Networks on Intel Loihi

TL;DR

This paper introduces NxTF, a specialized API and compiler for mapping deep spiking neural networks onto Intel Loihi neuromorphic hardware, enabling efficient deployment, model compression, and improved accuracy for event-driven processing tasks.

Contribution

NxTF provides a Keras-based programming interface and compiler optimized for deep SNNs on Loihi, supporting model distribution, compression, and achieving high resource utilization and accuracy.

Findings

Achieves 80% resource utilization across 16 Loihi chips for a 28-layer MobileNet.

Reports lowest error rate of 8.52% on CIFAR-10 using MobileNet on neuromorphic hardware.

Demonstrates effective model compression and energy-efficient processing.

Abstract

Spiking Neural Networks (SNNs) are a promising paradigm for efficient event-driven processing of spatio-temporally sparse data streams. SNNs have inspired the design and can take advantage of the emerging class of neuromorphic processors like Intel Loihi. These novel hardware architectures expose a variety of constraints that affect firmware, compiler and algorithm development alike. To enable rapid and flexible development of SNN algorithms on Loihi, we developed NxTF: a programming interface derived from Keras and compiler optimized for mapping deep convolutional SNNs to the multi-core Intel Loihi architecture. We evaluate NxTF on DNNs trained directly on spikes as well as models converted from traditional DNNs, processing both sparse event-based and dense frame-based data sets. Further, we assess the effectiveness of the compiler to distribute models across a large number of cores and to compress models by exploiting Loihi's weight sharing features. Finally, we evaluate model accuracy, energy and time to solution compared to other architectures. The compiler achieves near optimal resource utilization of 80% across 16 Loihi chips for a 28-layer, 4M parameter MobileNet model with input size 128x128. In addition, we report the lowest error rate of 8.52% for the CIFAR-10 dataset on neuromorphic hardware, using an off-the-shelf MobileNet.