Abstract
Spiking Neural Networks (SNNs) provide a brain-inspired, energy-efficient framework for machine learning by communicating via sparse binary spikes. However, their application to high-precision regression tasks, such as object detection, has been hindered by high latency and conversion errors. This paper presents a novel temporal-dependent Integrate-and-Fire (tdIF) neuron model and a delay-spike strategy designed to enable ultra-low-latency inference. By employing a dynamic threshold that implements binary temporal coding, the tdIF neuron represents complex values exponentially faster than traditional rate-coding models. Furthermore, the delay-spike strategy eliminates information loss from residual membrane potential. Experimental results on PASCAL VOC and MS COCO datasets demonstrate that the proposed method achieves state-of-the-art accuracy within 5 to 8 time-steps, offering a significant leap in SNN efficiency for real-time applications.
Introduction
Spiking Neural Networks (SNNs) are increasingly recognized for their potential in low-power edge computing. Unlike standard Artificial Neural Networks (ANNs), SNNs utilize discrete spikes, allowing hardware to replace energy-intensive multiplications with simple additions. Despite these advantages, SNNs face a significant challenge in regression tasks like object detection, where precise coordinate predictions are required. Traditional SNNs rely on rate coding, which necessitates hundreds or thousands of time-steps to achieve high precision, leading to unacceptable latency for real-time tasks. Additionally, the conversion from pre-trained ANNs to SNNs often introduces errors due to "residual potential"—the charge remaining in a neuron at the end of an inference window that fails to trigger a spike.
Methodology: tdIF and Delay-Spike Strategy
To address these issues, the authors propose three integrated innovations:
- Temporal-Dependent IF (tdIF) Neuron: The core of the approach is a neuron model with a dynamic threshold, $\theta(t) = 2^{T-t}$. This mechanism shifts the information encoding from spike frequency (rate coding) to spike timing (temporal coding). In this scheme, a spike at an earlier time-step carries a higher weight, effectively implementing a binary representation of values. This allows the network to achieve high precision with exponentially fewer time-steps; for instance, 3 time-steps in a tdIF model can represent as much information as 7 time-steps in a standard IF model.
- Delay-Spike Strategy: To mitigate errors from residual membrane potential, a two-stage firing process is introduced. Neurons first accumulate potential without firing, then release all accumulated information as a burst of spikes in the second stage. This ensures that no "leftover" energy is lost, preserving the mathematical integrity of the ANN-to-SNN conversion.
- Quant-Clip Activation: During the ANN training phase, a specialized activation function is used to simulate the quantization effects of the target SNN. This ensures that the weights optimized during training are perfectly suited for the spiking domain.
Experimental Results
The proposed model was evaluated on several demanding computer vision benchmarks, including PASCAL VOC and MS COCO for object detection, and TuSimple and CULane for lane detection.
The results demonstrate a dramatic reduction in latency without sacrificing accuracy. On the PASCAL VOC dataset, the model achieved a mean Average Precision (mAP) of 74.41% using a YOLOv3-ResNet34 backbone in just 5 to 8 time-steps. For comparison, previous state-of-the-art models like Spiking-YOLO required up to 2,000 steps to reach similar performance levels. In lane detection tasks on the TuSimple dataset, the model reached 95.48% accuracy within 5 time-steps, proving its suitability for high-speed automotive safety applications.
Discussion and Conclusion
The shift from counting spikes to timing spikes represents a fundamental advancement in SNN design. By assigning different informational values to spikes based on their temporal position, the tdIF model maximizes the information density per time-step. While the tdIF neuron is slightly more complex than standard models—requiring hardware that supports bit-shifting or dynamic thresholding—the trade-off is a massive gain in speed and accuracy. Future work will focus on optimizing hardware implementations for these dynamic thresholds and exploring applications in sub-millisecond real-time robotic control where ultra-low latency is critical.
Summary updated