Weight and Activation Quantization

 The experiments showed that the tradeoff between accuracy and

quantization differs in each layer. We call this layer sensitivity.
 For example, middle layers are much more robust to quantization
whilst also housing majority of the weights
 As shown in Figure, majority of the MAC operations are conducted
in the second, third and fourth convolutional layers.
 Using mixed2 scheme, more than 3 times saves is achieved at the
MAC operations. Hence, the energy consumption and latency
numbers can also be significantly reduced with only sacrificing
0.2% accuracy.
 In the next Figure , the distribution of weights per layer is shown.
 It is clear that the majority of weights reside in the first dense
layer, followed by the fourth convolutional layer.
 These are also some of the least sensitive layers. Meaning, by
deeply quantizing these layers, majority of the memory space
can be saved