Table 1.

Layer precision for each model version (FP = floating-point; B = binary; Q = quantized).

Model version	FP model	B+FP model	B+Q model
Most conv layers	FP	Binary	Binary
First layer	FP	FP	4-bit input + 4-bit weight
Skip connections	FP	FP	4-bit/6-bit with scalable range per connection
BatchNorm	FP	FP	Binary shift + 4-bit bias
Activation function	ReLU	PReLU with FP $\beta$, $\gamma$ & $\eta$	PReLU with 2-bit $\beta$ and 4-bit $\gamma$ & $\eta$
RSign	–	FP threshold ($\alpha$)	4-bit threshold ($\alpha$)