Presentation
LEAF: An Adaptation Framework against Noisy Data on Edge through Ultra Low-Cost Training
DescriptionIn real-world neural network deployments, incoming data often contains noise and imperfections. Retraining on resource-constrained edge devices becomes essential to maintain performance. To tackle this challenge, we introduce LEAF, a hardware-efficient framework designed for adapting to degraded images. By analyzing neural network behavior on degraded images, we propose two techniques: 1) Selective Experience Replay for skipping unimportant images, reducing computation, and 2) Pseudo Noise Dithering for extremely low precision (3 or 4-bit) gradient quantization, enabling nearly full-integer training. Extensive experiments on CIFAR10 and Tiny ImageNet datasets, with various image degradations, demonstrate LEAF's ultra-low cost with minimal accuracy loss.
Event Type
Research Manuscript
TimeTuesday, June 2510:45am - 11:00am PDT
Location3001, 3rd Floor
AI
AI/ML Algorithms