Table 3.

CIFAR10 network topologies for Conversion training.

Model	Configuration	BackRes
VGG7	Input–Conv1(3,64,3 × 3/1)–Conv2(64,64,3 × 3/1)–	Not applicable
	–Conv3(64,64,3 × 3/1)– Conv4(64,64,3 × 3/1)–
	–Conv5(64,64,3 × 3/1)–Pool(2x2/2)–Pool(2 × 2/2)–
	–Pool(2 × 2/2)–FC1(2048,512)–FC2(512,10)
VGG2x4	Input–Conv1(3,64,3 × 3/1)–$\text{Conv2}(\mathrm{64,64,3}\times \mathrm{3/1})$–	$\text{[Conv2]}$ repeated 4 times
	–$\text{Conv2}(\mathrm{64,64,3}\times \mathrm{3/1})$–$\text{Conv2}(\mathrm{64,64,3}\times \mathrm{3/1})$–
	–$\text{Conv2}(\mathrm{64,64,3}\times \mathrm{3/1})$–Pool(2 × 2/2)–Pool(2 × 2/2)–
	–Pool(2 × 2/2)–FC1(2048,512)–FC2(512,10)
VGG3x2-v1	Input–Conv1(3,64,3 × 3/1)–$\text{Conv2}(\mathrm{64,64,3}\times \mathrm{3/1})$–	$\text{[Conv2–Conv3]}$ repeated 2 times
	–$\text{Conv3}(\mathrm{64,64,3}\times \mathrm{3/1})$–$\text{Conv2}(\mathrm{64,64,3}\times \mathrm{3/1})$–
	–$\text{Conv3}(\mathrm{64,64,3}\times \mathrm{3/1})$–Pool(2 × 2/2)–Pool(2 × 2/2)–
	–Pool(2 × 2/2)–FC1(2048,512)–FC2(512,10)

ConvN(I,O,k × k/s) denotes N^th convolutional layer with I input channels, O output channels, kernel of size k × k with stride s. Pool(p × p/s_p) denotes average pooling layer with pooling window size p × p and pooling stride s_p. FC(X,Y) denote a fully-connected layer with X input nodes and Y output nodes. Layers with BackRes connections and repeated computations have been highlighted in red.