Abstract
Spiking Neural networks (SNN) have emerged as an attractive spatio-temporal computing paradigm for a wide range of low-power vision tasks. However, state-of-the-art (SOTA) SNN models either incur multiple time steps which hinder their deployment in real-time use cases or increase the training complexity significantly. To mitigate this concern, we present a training framework (from scratch) for one-time-step SNNs that uses a novel variant of the recently proposed Hoyer regularizer. We estimate the threshold of each SNN layer as the Hoyer extremum of a clipped version of its activation map, where the clipping threshold is trained using gradient descent with our Hoyer regularizer. This approach not only downscales the value of the trainable threshold, thereby emitting a large number of spikes for weight update with a limited number of iterations (due to only one time step) but also shifts the membrane potential values away from the threshold, thereby mitigating the effect of noise that can degrade the SNN accuracy. Our approach outperforms existing spiking, binary, and adder neural networks in terms of the accuracy-FLOPs trade-off for complex image recognition tasks. Downstream experiments on object detection also demonstrate the efficacy of our approach.
Introduction
Spiking Neural Networks (SNNs) are often hailed as the future of low-power artificial intelligence due to their ability to mimic the brain's event-driven communication. By utilizing discrete spikes instead of continuous values, SNNs can significantly reduce energy consumption. However, traditional SNNs face a "latency tax," typically requiring multiple time-steps (often 5 to 100) to achieve high accuracy. This requirement introduces significant latency and complicates training, as gradients must be integrated over time.
Recent research from the University of Southern California introduces a framework that enables high-accuracy SNNs to operate in just a single time-step while being trained from scratch, bypassing the need for expensive ANN-to-SNN conversion.
The Hoyer Spike Framework
The core innovation of this work is the Hoyer Spike Layer. In a single-time-step environment, the choice of firing threshold is critical. If the threshold is too high, the network remains silent; if too low, it produces noise. The authors address this by using a modified Hoyer regularizer—a sparsity measure based on the ratio of $L_1$ and $L_2$ norms—to dynamically determine neuron thresholds.
As shown in Figure 1, the threshold for each layer is estimated as the "Hoyer extremum" of clipped membrane potentials. The clipping threshold itself is a trainable parameter optimized via gradient descent. This mechanism serves two purposes: it downscales the threshold to ensure sufficient spiking for weight updates in a single step, and it shifts the distribution of membrane potentials away from the threshold. This "distribution shifting" increases the signal-to-noise ratio, making the network more robust to perturbations.
The architectural implementation, illustrated in Figure 2, replaces standard Leaky Integrate-and-Fire (LIF) blocks with the proposed Hoyer Spike blocks. Unlike existing methods that rely on average pooling and LIF spikes, the proposed block integrates Max Pooling and Batch Normalization (BN) before the Hoyer Spike layer, followed by Dropout and Convolution. This structure allows the model to be trained directly using standard Kaiming initialization.
Experimental Results
The researchers evaluated the framework across several benchmarks, including CIFAR10, ImageNet-1K, and object detection tasks (PASCAL VOC).
- Accuracy: On ImageNet, the model achieved a top-1 accuracy of 68.00%, which is competitive with non-spiking counterparts while maintaining extreme efficiency. On CIFAR10, it reached 93.44% accuracy.
- Efficiency: The framework achieved a 5.5x reduction in FLOPs compared to Binary Neural Networks (BNNs) and Adder Neural Networks (AddNNs). Furthermore, the single-time-step nature resulted in a 19x reduction in training time compared to iterative SNN methods.
- Energy Consumption: Theoretical estimates on systolic array accelerators suggest a 22.5x to 31.4x reduction in total energy compared to standard deep neural networks.
Conclusion and Future Work
By leveraging the Hoyer regularizer to optimize neuron thresholds, this framework demonstrates that ultra-low latency SNNs can achieve high performance without the complexity of multi-step unrolling or pre-training. While a small accuracy gap remains in complex tasks like object detection (approximately 5.8–6.8% mAP), the massive gains in training speed and energy efficiency mark a significant step toward real-time, low-power neuromorphic hardware deployment. Future work will focus on narrowing the detection gap and exploring the adversarial robustness of these discrete spiking systems.
Summary updated