60

Algorithms for Binary Neural Networks

FIGURE 3.16

Training and testing curves of PCNN-22 when λ=0 and 1e −4, which shows that the

projection affects little on the convergence.

3.6

RBCN: Rectified Binary Convolutional Networks with Gener-

ative Adversarial Learning

Quantization approaches represent network weights and activations with fixed-point integers

of low bit width, allowing computation with efficient bitwise operations. Binarization [199,

159] is an extreme quantization approach where both weights and activations are +1 or −1,

represented by a single bit. This chapter designs highly compact binary neural networks

(BNNs) from the perspective of quantization and network pruning.







FIGURE 3.17

Illustration of binary kernels D^l

i ^{(first row), feature maps produced by Dl}

i ^{(second row),}

and corresponding feature maps after binarization (third row) when J=4. This confirms

the diversity in PCNNs.