Introduction
Autonomous Driving (AD) systems require a delicate balance of high accuracy, real-time processing, and low power consumption. Spiking Neural Networks (SNNs) have emerged as a promising solution for these requirements. Unlike traditional artificial neural networks, SNNs mimic the brain's energy efficiency by processing information through discrete spikes. They are particularly well-suited for "event-based" cameras, which record changes in pixel brightness as they occur rather than capturing full frames at fixed intervals. This synergy allows for high-speed perception with minimal data redundancy.
Despite their potential, most SNN research has focused on designing novel architectures. There has been a notable lack of systematic study regarding how SNN hyperparameters—the internal settings that govern learning—impact performance when dealing with the noisy, sparse data typical of automotive environments. Researchers from NYU Abu Dhabi have addressed this gap by developing a systematic methodology to optimize SNN parameters specifically for event-based automotive vision, demonstrating that fine-tuning existing architectures can yield significant gains in both accuracy and efficiency.
Methodology
The researchers proposed a three-step framework to move beyond trial-and-error parameter selection:
- Identification: The study identified four critical parameters influencing the Spatio-Temporal Back-Propagation (STBP) learning rule: Batch Size ($B$), Learning Rate ($lr$), Neuron Threshold Potential ($V_{th}$), and Weight Decay ($w_{decay}$).
- Investigation: Using a "grid search-lite" approach, the team varied one parameter at a time while keeping others constant. This allowed them to observe how each setting affected the "transition region" (early training) and the "stable region" (final convergence) of the learning curve.
- Enhancement: By aggregating the effective values discovered during the investigation, the researchers created an optimized SNN model configuration.
The underlying architecture for this study was a 7-layer SNN, incorporating convolutional and pooling layers, trained using the STBP rule to account for both spatial patterns and temporal spike dynamics.
Experimental Setup and Dataset
The methodology was validated using the Prophesee NCARS Dataset, a benchmark for event-based automotive vision. The dataset contains approximately 24,000 samples of 100ms recordings, focusing on a binary classification task: distinguishing between a "car" and "background."
Unlike standard image datasets, NCARS data consists of a stream of events $(x, y, t, p)$, representing pixel coordinates, timestamps, and polarity changes. This format requires the SNN to process temporal sequences effectively, making the choice of hyperparameters like threshold potential and learning rate critical for success.
Key Insights on Parameter Impact
The investigation revealed several counter-intuitive findings that challenge standard deep learning conventions:
- The Advantage of Small Batches: In traditional deep learning, larger batches are often used to stabilize training. However, for the NCARS dataset, the authors found that smaller batch sizes ($B=20$) performed better. Given the binary nature of the task, smaller batches provided finer-grained updates that helped the network avoid overfitting to the background noise.
- The Threshold "Goldilocks" Zone: The threshold potential ($V_{th}$) determines the voltage at which a neuron fires. If set too high (e.g., 0.7), neurons fire too infrequently, starving the network of information. The researchers identified $V_{th} = 0.5$ as the optimal value for maintaining healthy spiking activity across the layers.
- Weight Decay Sensitivity: While weight decay is a standard regularization technique to prevent overfitting, it proved counter-productive for this SNN task. Any value of $w_{decay} > 0$ caused the network to "unlearn" critical temporal information over time, leading to a significant drop in accuracy.
- Accelerated Convergence: By optimizing the learning rate to $7.5e-3$, the network reached its stable performance region much faster. The optimized model achieved 85% accuracy in just 66 epochs, compared to the 125 epochs required by the baseline model.
Results and Performance
Applying this systematic methodology led to substantial improvements over the baseline CarSNN model:
- State-of-the-Art Accuracy: The optimized model achieved a peak accuracy of 86.64% on the NCARS dataset.
- Training Efficiency: The model reached "iso-accuracy" (performance comparable to the baseline) 1.9x faster.
- Robustness: The optimized parameters resulted in a smoother learning curve with a standard deviation of less than 0.5% in the stable region, indicating a highly predictable and robust training process.
Conclusion and Future Work
This research provides a roadmap for SNN parameter enhancement, proving that systematic tuning is as vital as architectural innovation. While the study focused on binary classification, it sets the stage for more complex automotive tasks.
Future work will look toward multi-object detection and semantic segmentation. Additionally, while this study utilized high-end GPUs (Nvidia RTX 6000), the researchers aim to explore how these parameter choices translate to neuromorphic hardware, such as Intel’s Loihi, where physical constraints may further influence optimal hyperparameter settings. Ultimately, this methodology offers a foundational step toward making energy-efficient SNNs a reality for next-generation autonomous vehicles.
Summary updated