FastSpiker: Enabling Fast Training for Spiking Neural Networks on Event-based Data through Learning Rate Enhancements for Autonomous Embedded Systems

Iqra Bano, Rachmad Vidya Wicaksana Putra, Alberto Marchisio, Muhammad Shafique

Abstract

Autonomous embedded systems, such as drones and robotic agents, require high-efficiency intelligence to operate within strict power constraints. Spiking Neural Networks (SNNs) paired with event-based sensors offer a biologically inspired solution for low-power processing. However, the training phase for SNNs remains computationally expensive and slow, hindering the potential for online learning and sustainable development. This paper introduces FastSpiker, a methodology that optimizes Learning Rate (LR) policies to accelerate SNN training. By systematically evaluating LR ranges, scheduling policies, and hyperparameter interactions, FastSpiker achieves up to a 10.5x speedup in training time and an 88.39% reduction in carbon emissions on automotive event-based datasets without compromising accuracy.

Introduction

As autonomous systems move toward edge deployment, the demand for energy-efficient AI has never been higher. Spiking Neural Networks (SNNs) are uniquely suited for these applications because they process information using sparse, discrete spikes, mimicking the energy efficiency of the biological brain. When coupled with event-based sensors like Dynamic Vision Sensors (DVS), which only record changes in pixel intensity, SNNs provide a high-temporal-resolution, low-power alternative to traditional frame-based computer vision.

Despite these advantages, SNNs suffer from a significant "slow-to-learn" bottleneck. Training state-of-the-art SNN models on event-based data can take several hours on high-end hardware, leading to substantial energy consumption and a high carbon footprint. This latency also prevents "online training," where an autonomous system adapts to new environments in real-time. FastSpiker addresses this challenge by focusing on the optimization of learning rate schedules, a critical yet often overlooked component of SNN training efficiency.

The FastSpiker Methodology

The FastSpiker framework moves away from static training schedules in favor of a dynamic, three-step optimization process designed to find the fastest path to model stability.

1. Determining the Effective LR Range

The first step involves a systematic "sweep" of learning rate values, typically ranging from $1e-6$ to $1e-1$. By monitoring the loss landscape, researchers identify the "sweet spot"—the range where the network demonstrates consistent learning without diverging. For the automotive datasets tested, this range was identified between $1e-5$ and $1e-2$.

2. Evaluating LR Policies

FastSpiker evaluates six distinct LR scheduling strategies to determine which facilitates the fastest convergence:

  • Decreasing Step & Exponential Decay: Gradually reducing the LR to fine-tune weights.
  • One-Cycle & Cyclical LR: Oscillating the LR to navigate complex loss landscapes.
  • Decreasing Cyclical & Warm Restarts: Periodically resetting the LR to high values to escape local minima.

3. Parameter Exploration and Stability

Beyond the LR schedule, FastSpiker explores the interaction between the LR and other critical hyperparameters, specifically Batch Size ($B$) and the Neuron’s Membrane Threshold ($V_{th}$). To objectively measure training speed, the authors define a "stability threshold": training is considered complete when the model's accuracy fluctuates by less than 1% of the standard deviation over the final 10 epochs. This allows for an "early exit," preventing unnecessary computation once the model has converged.

Experimental Evaluation

The methodology was validated using the Prophesees’ NCARS dataset, a standard benchmark for event-based automotive tasks. The dataset consists of approximately 24,000 samples of 100ms duration, where the task is binary classification (distinguishing between "car" and "background"). Unlike traditional video, this data is a sparse sequence of events containing spatial coordinates, timestamps, and polarity.

Key Results and Insights

Training Acceleration and Sustainability

The FastSpiker methodology yielded dramatic improvements over state-of-the-art (SOTA) baselines:

  • Speedup: The best-performing configuration (Exponential Decay with optimized batch size) reached stability 10.5x faster than the baseline.
  • Efficiency: While the baseline required 77 epochs to stabilize, FastSpiker reached the same performance in just 19 epochs.
  • Environmental Impact: The reduction in training time translated directly to an 88.39% decrease in estimated carbon emissions, supporting the goals of "Green AI."

Policy Performance

  • The "Start High" Advantage: Policies like Exponential Decay that begin with a high LR allow the network to find a global minimum rapidly. In binary tasks like NCARS, this aggressive start does not trigger the instability often seen in more complex multi-class problems.
  • Warm Restarts: While slightly slower to stabilize than Exponential Decay, the Warm Restarts policy proved superior for reaching peak accuracy by effectively escaping local minima during the training process.

Discussion and Limitations

While FastSpiker demonstrates a clear path toward efficient SNN training, several areas for future research remain. The current evaluation focused on binary classification; further testing is required to see if these LR enhancements scale to multi-class object detection or complex tracking tasks. Additionally, while the training was optimized on Nvidia RTX 6000 GPUs, the ultimate goal is to implement these policies directly on neuromorphic hardware (such as Intel’s Loihi) to enable true on-device learning.

Conclusion

FastSpiker provides a robust framework for accelerating the training of Spiking Neural Networks on event-based data. By optimizing learning rate policies and defining clear stability thresholds, the methodology significantly reduces the time and energy required to develop intelligent autonomous systems. This work represents a vital step toward making SNNs a practical and sustainable choice for the next generation of edge-based AI.

Summary updated

Abstract

Autonomous embedded systems (e.g., robots) typically necessitate intelligent computation with low power/energy processing for completing their tasks. Such requirements can be fulfilled by embodied neuromorphic intelligence with spiking neural networks (SNNs) because of their high learning quality (e.g., accuracy) and sparse computation. Here, the employment of event-based data is preferred to ensure seamless connectivity between input and processing parts. However, state-of-the-art SNNs still face a long training time to achieve high accuracy, thereby incurring high energy consumption and producing a high rate of carbon emission. Toward this, we propose FastSpiker, a novel methodology that enables fast SNN training on event-based data through learning rate enhancements targeting autonomous embedded systems. In FastSpiker, we first investigate the impact of different learning rate policies and their values, then select the ones that quickly offer high accuracy. Afterward, we explore different settings for the selected learning rate policies to find the appropriate policies through a statistical-based decision. Experimental results show that our FastSpiker offers up to 10.5x faster training time and up to 88.39% lower carbon emission to achieve higher or comparable accuracy to the state-of-the-art on the event-based automotive dataset (i.e., NCARS). In this manner, our FastSpiker methodology paves the way for green and sustainable computing in realizing embodied neuromorphic intelligence for autonomous embedded systems.
BibTeX
@inproceedings{bano2024fastspiker,
  author = {Iqra Bano and Rachmad Vidya Wicaksana Putra and Alberto Marchisio and Muhammad Shafique},
  title = {{FastSpiker: Enabling Fast Training for Spiking Neural Networks on Event-based Data through Learning Rate Enhancements for Autonomous Embedded Systems}},
  booktitle = {ICARCV},
  year = {2024},
  doi = {10.1109/icarcv63323.2024.10821701},
}
Figure