Energy-efficient traffic sign recognition using directly trained spiking neural networks and population decoding

Area of Science:

• Computer Vision
• Artificial Intelligence
• Neuromorphic Computing

Background:

• Traffic sign recognition is crucial for autonomous driving, demanding high accuracy under strict latency and energy constraints.
• Convolutional neural networks (CNNs) are effective but computationally intensive for embedded systems.
• Spiking convolutional neural networks (SCNNs) offer energy efficiency via sparse, event-driven computation but struggle with accuracy at low latencies.

Purpose of the Study:

• To enhance SCNN accuracy and efficiency for traffic sign recognition under strict runtime constraints.
• To evaluate the impact of a neural population decoding layer on SCNN performance.
• To compare directly trained SCNNs with and without population decoding against a CNN baseline.

Main Methods:

• Directly trained SCNNs with and without a population decoding layer were implemented.
• Performance was evaluated on the German Traffic Sign Recognition Benchmark (GTSRB).
• Accuracy, inference time, and energy efficiency were measured and compared to a CNN baseline.

Main Results:

• The best SCNN achieved 98.85% accuracy at 30 time steps, surpassing the CNN baseline (98.38%).
• Population decoding improved SCNN accuracy to 98.31% at a single time step.
• SCNNs demonstrated over 14 times higher energy efficiency than CNNs at one time step.

Conclusions:

• Directly trained SCNNs can outperform comparable CNNs in traffic sign recognition.
• SCNNs offer flexible trade-offs between accuracy, inference time, and energy efficiency.
• Population decoding is particularly beneficial for SCNNs operating under strict latency constraints.