Deep learning for computer vision at Lighthouse AI
Despite of the success of deep networks, they could be easily fooled by few pixels of noise so as to output incorrect answers. Our feedforward model utilizes uncertainty information and achieves high robustness against strong noise on a large-scale dataset.
Since deep networks contain massive and redundant parameters, finding a compact representation is a key to save runtime and memory. Our model based on filter separation can substitute for 3D filters without accuracy loss and provides 2x speed-up with 90% parameter reduction.
Deep networks are state-of-the-art models for visual understanding but require off-site server for processing due to heavy computational cost. Our accelerator on a mobile coprocessor is capable of a peak performance of 240 G-ops/s while consuming less than 4 W.