Val accuracy is so bad?

[MrLinNing]

How to make the network more accuracy?

[layog]

It's almost impossible to reach the accuracy of the normal convolutional layers. That being said, I was able to achieve upto 92% accuracy on MNIST (using my own trained network, though I believe you can achieve higher accuracy using a pretrained network). You can try to increase the number of binary layers and ApproxConv layers and check if that gives some boost to the accuracy.