REMODEL: Rethinking Deep CNN Models to Detect and Count on a NeuroSynaptic System

Area of Science:

• Computer Vision
• Neuromorphic Computing
• Artificial Intelligence

Background:

• The increasing demand for real-time image analysis necessitates low-power, efficient hardware solutions.
• Traditional GPUs consume significant power, limiting their deployment in edge computing scenarios.
• Neuromorphic systems offer a promising alternative for energy-efficient AI tasks.

Purpose of the Study:

• To analyze the detection and counting of cars using the low-power IBM TrueNorth Neurosynaptic System.
• To identify and address architectural bottlenecks in the TrueNorth system for deploying larger neural networks.
• To compare the performance and power efficiency of the TrueNorth system against GPU-based implementations.

Main Methods:

• A trained Convolutional Neural Network (CNN) for image analysis was deployed on the IBM NS16e system using the EEDN training framework.
• Experiments were conducted using a publicly available dataset of overhead car imagery.
• The TrueNorth system's performance was benchmarked against Caffe-based neural network implementations on a Titan-X GPU.

Main Results:

• The TrueNorth system achieved 97.60% accuracy in car detection.
• The system demonstrated 69.04% accuracy in counting cars within a +/- 2 error margin.
• Car detection and counting accuracy were comparable to high-precision networks like AlexNet, GoogLeNet, and ResCeption.

Conclusions:

• The IBM TrueNorth Neurosynaptic System offers a viable, low-power solution for car detection and counting tasks.
• The system provides a significant improvement in power consumption compared to GPU-based solutions.
• Adaptations to CNN models can circumvent architectural limitations of the TrueNorth system, enabling efficient deployment of complex AI models.