Xiangxiang Xu
Aug 22, 2017 · 1 min read

Hi Claude, thank you for sharing! I am also reading Deep Learning and trying to implement this toy model, and have two questions:

  1. As the neural network is quite sensitive to initial weights, I wonder whether you successfully converge to solve the XOR problem each time. In other words, have you further investigated the convergence ratio in batch running with different initial weights?
  2. In your cross-entropy loss function case, you actually run sigmoid twice: one is calling sigmoid in evaluate y_estimated, the other is calling sigmoid_cross_entropy_with_logits. I think directly using sigmoid_cross_entropy_with_logits(tf.add(tf.matmul(h,w),b)) is enough.

)

    Xiangxiang Xu

    Written by

    https://xiangxiangxu.com/