- trevi_
- Korea
- isuu.jeong@gmail.com
- https://github.com/ijgit
Binarized Neural Networks: Training Neural Networks with Weights and Activations Constrained to +1 or -1 (1)
Abstract BNNs (Binarized Neural Networks) neural networks with binary wrights and activations at run-time at trining-time the binary weights and activations are userd for computing the gradients 방법 forward pass 동안, 메모리 사이즈와 접근을 줄이고, 대부분의 산술 연산을 비트 연산으로 대체함으로써 전력 효율성을 개선할 수 있다. 검증 two sets of experiments on Torch7 and Theano frameworks 검증 결과 MNIST, CIFAR-10, SVHN datasets 에서 매우 좋은 결과를 얻었다. binary..