Lightweight Local Feature Restoration for Reliable Energy-Efficient Feature Extraction

This work proposes Local Feature Restoration (LFR) for robust, energy-efficient feature extraction and data processing at the sensor node. Analog Feature Extraction (AFE) reduces digitization and communication overheads by applying downsizing convolution to compress incoming data. However, analog computation introduces multiple noise sources—including linearity errors, digitization errors, and device-level variations in novel technologies—that degrade overall performance. Local Feature Restoration (LFR) mitigates this accuracy loss by restoring features from noise with minimal overhead. With the addition of DACs, LFR can enable all data processing to finish at the sensor node. On ImageNet200 and CIFAR10, feature extraction with LFR achieves 5X reduction in energy while maintaining accuracy within 0.04-0.67% of ideal/baseline models with only 1-7% area/energy overhead for LFR. In-sensor data processing delivers 40X energy savings with just 3.15% accuracy drop on CIFAR10 and under 3% LFR energy/area overhead. Without AFE, LFR restores accuracy to within 0.7% of the ideal quantized model on ImageNet200. Through localized error correction, LFR enables reliable, energy-efficient feature extraction and on-chip data processing for edge and sensor computing applications.