Quantized Convolutional Neural Networks for Mobile Devices

cvpr 2016

1. Problem

- CNN needs high performance hardware which prohibits their further extension (e.g. mobile service)

2. Solution

- Simultaneously speed-up the computation and reduce the storage and memory overhead of CNN models > Quantized CNN

3. How

- Quantized CNN
  + Both filter kernels in convolutional layers and weighting matrices in fully connected layers are quantized, aiming at minimizing the estimation error of each layer's response.

4. Performance

- ILSVRC-12
  + 4~6 X speed-up and 15 ~20 X compression with merely one percentage loss of classification accuracy
- even mobile devices can accurately classify images within one second.