PaperlayerView exact revision on arXiv

Source-linked AI summary

Spiking-YOLO: Spiking Neural Network for Energy-Efficient Object Detection

Seijoon Kim, Seongsik Park, Byunggook Na, Sungroh Yoon

arXiv:1903.06530v2cs.CVcs.LGstat.ML

TL;DR

SNNs promise energy-efficient computation but have mainly been applied to simpler classification tasks and face training challenges. This paper introduces Spiking-YOLO for object detection using channel-wise normalization and signed neurons with imbalanced thresholds. It reports comparable detection performance to Tiny YOLO, substantially lower energy use, and faster convergence than prior conversion methods.

  • Problem

    SNN applications have largely been limited to relatively simple classification tasks, while object detection requires precise bounding-box regression.

  • Method

    The paper introduces Spiking-YOLO with channel-wise normalization and signed neurons with imbalanced thresholds for deep-SNN object detection.

  • Results

    Spiking-YOLO achieves up to 98% of Tiny YOLO’s performance on PASCAL VOC and MS COCO, while its full method reaches 51.83% mAP on PASCAL VOC and 25.66% on MS COCO.

  • Takeaways & Limitations

    Spiking-YOLO demonstrates that deep SNNs can perform object detection on non-trivial datasets with comparable DNN performance.

Abstract

from arXiv · show

Over the past decade, deep neural networks (DNNs) have demonstrated remarkable performance in a variety of applications. As we try to solve more advanced problems, increasing demands for computing and power resources has become inevitable. Spiking neural networks (SNNs) have attracted widespread interest as the third-generation of neural networks due to their event-driven and low-powered nature. SNNs, however, are difficult to train, mainly owing to their complex dynamics of neurons and non-differentiable spike operations. Furthermore, their applications have been limited to relatively simple tasks such as image classification. In this study, we investigate the performance degradation of SNNs in a more challenging regression problem (i.e., object detection). Through our in-depth analysis, we introduce two novel methods: channel-wise normalization and signed neuron with imbalanced threshold, both of which provide fast and accurate information transmission for deep SNNs. Consequently, we present a first spiked-based object detection model, called Spiking-YOLO. Our experiments show that Spiking-YOLO achieves remarkable results that are comparable (up to 98%) to those of Tiny YOLO on non-trivial datasets, PASCAL VOC and MS COCO. Furthermore, Spiking-YOLO on a neuromorphic chip consumes approximately 280 times less energy than Tiny YOLO and converges 2.3 to 4 times faster than previous SNN conversion methods.

Introduction

The paper targets the high computational demands and limited task scope of SNNs by applying them to object detection. It introduces Spiking-YOLO with channel-wise normalization and signed neurons to achieve accurate, efficient detection.

  • SNNs offer sparse, low-powered computation but remain difficult to train because of complex neuron dynamics and non-differentiable spike operations.
  • Object detection extends SNN evaluation beyond classification by requiring recognition of multiple overlapping objects and precise bounding-box regression.
  • Spiking-YOLO combines channel-wise normalization with signed neurons using imbalanced thresholds to improve information transmission in deep SNNs.
  • Spiking-YOLO provides the first deep-SNN object detector with comparable performance to DNNs on PASCAL VOC and MS COCO.
  • The proposed signed-neuron method implements leaky-ReLU efficiently in SNNs and can be deployed on neuromorphic chips with minimal overhead.

Related work

Prior work established event-driven SNN computation, training approaches, and DNN-to-SNN conversion, but mainly addressed classification. Object detection combines classification with precise coordinate regression, motivating the use of fast one-stage architectures such as Tiny YOLO.

  • Integrate-and-fire neurons accumulate incoming spikes in membrane potentials and emit spikes when the integrated value exceeds a threshold.
  • SNNs perform energy-efficient event-driven operations but are difficult to train because spike operations are non-differentiable.
  • SNN training includes unsupervised STDP and supervised gradient-based methods, with supervised learning generally achieving better performance.
  • DNN-to-SNN conversion transfers pretrained DNN parameters, but previous conversion work was largely limited to image classification.
  • Object detection requires both object classification and precise prediction of bounding-box coordinates and dimensions.
  • R-CNN-based detectors use multiple stages and suffer slow inference, whereas one-stage methods such as YOLO prioritize real-time speed.

Methods

The methods address performance degradation in deep SNN object detection by improving activation scaling and implementing leaky-ReLU with signed neurons. Channel-wise normalization raises firing rates and supports faster, more accurate information transmission, while imbalanced thresholds enable efficient signed-neuron dynamics.

  • Motivation: Object detection in deep SNNs suffers from performance degradation because low firing rates and inefficient leaky-ReLU implementation impair precise regression.Object detection requires high numerical precision for bounding-box coordinates, making information loss especially problematic.
  • Channel-wise normalization: Layer-wise normalization produces highly uneven channel activations, with some Conv1 channels near 1 and others near 0.Tiny YOLO contains eight convolutional layers, and the channel deviations recur across layers.
  • Channel-wise normalization: Channel-wise normalization scales weights using each channel’s 99.9th-percentile maximum activation instead of a layer-wide value.The method is designed to eliminate extremely small activations and preserve activation information across channels.
  • Channel-wise normalization: Channel-wise normalization produces higher firing rates across most channels, including a 20-times-higher rate than layer-norm in Conv1 channel 2.Its firing-rate distributions and raster plots show more neurons firing regularly, supporting accurate transmission in fewer time steps.
  • Signed neuron: A signed neuron with imbalanced thresholds implements the positive and negative regions of leaky-ReLU in an SNN.For α = 0.1, the positive threshold is 1V while the negative threshold is −10V, requiring ten times more integration for negative activations.

Evaluation

Spiking-YOLO is evaluated on PASCAL VOC and MS COCO using Tiny YOLO as its base, with ablations of normalization, signed neurons, and output decoding. The proposed configuration approaches Tiny YOLO accuracy while improving convergence and energy efficiency.

  • Experimental setup: Tiny YOLO provides the real-time object-detection base for experiments on PASCAL VOC and MS COCO.The implementation includes max-pooling and batch-normalization in SNNs and uses NVIDIA Tesla V100 GPUs for simulation.
  • Detection accuracy: 51.83% and 25.66% mAP are achieved on PASCAL VOC and MS COCO, respectively, with channel-norm and signed neurons with IBT.Tiny YOLO targets 53.01% and 26.24% mAP on the two datasets.
  • Detection accuracy: 53.01% versus 48.94% mAP shows channel-norm outperforming layer-norm on PASCAL VOC.Channel-norm reaches layer-norm’s maximum mAP in approximately 3,500 time steps, or 2.3x faster.
  • Convergence: 4x faster convergence is observed for channel-norm than layer-norm on MS COCO.The passage reports similar convergence advantages on MS COCO and identifies channel-norm as faster on both datasets.
  • Ablation: 6.87% and 2.82% mAP without the proposed methods indicate failure to detect objects on PASCAL VOC and MS COCO.Channel-norm alone reaches only approximately 7.31% and 3.02%, while signed neuron with IBT is retained as the default.
  • Decoding: Vmem-based decoding outperforms spike-count decoding and converges faster with channel-norm.Spike-count decoding rounds the remainder of Vmem / Vth, whereas Vmem-based decoding preserves more precise output information.
  • Energy efficiency: 280 times less energy is consumed by Spiking-YOLO than Tiny YOLO when run on TrueNorth.The energy comparison considers Tiny YOLO on a GPU and Spiking-YOLO on a neuromorphic chip, including MAC and AC operation costs.
  • Energy efficiency: 4 times less energy is consumed by channel-norm than layer-norm because their power consumption is similar and channel-norm converges faster.The comparison is reported for Spiking-YOLO configurations with similar power consumption.

Conclusion

The paper presents Spiking-YOLO as the first SNN model to perform object detection with results comparable to original DNNs on PASCAL VOC and MS COCO. It introduces two novel methods and frames the work as a first step toward advanced machine-learning problems in deep SNNs.

  • Conclusion: Spiking-YOLO is presented as the first SNN model to successfully perform object detection on PASCAL VOC and MS COCO with DNN-comparable results.The conclusion describes the work as a first step toward more advanced machine-learning problems in deep SNNs.
  • Conclusion: Two novel methods are proposed as part of Spiking-YOLO’s object-detection approach.The supplied conclusion identifies the methods as part of the presented model but does not name them.
Loading 1903.06530v2…