• Mengqi

    Dear Eric,

    Thanks for your sharing. I have been wondering why the output of this autoencoder remain positi(ve even the input signal is in the range of (-1,1). Any comments?


  • Does not matter

    Mengqi, the input “rescaled from [-1,1] to [0.1,0.9]”. See the bottom of sampleIMAGES.m.

  • ali

    Dear Eric,
    could you please set up a skype call?

  • bubo

    Dear Eric,

    I find your post very useful, especially vectorisation of the cost func. One thing that I noticed after trying your code is that the output is actually scaled for some reason. I mean, you normalize the input to have it between [0.1,0.9]. However after training I tried to reconstruct the first patch (by passing it through the autoencoder), and I got ranges [0.48, 0.52] instead of [0.25 0.9]. After normalizing it to the original range it looks ok, but I don’t understand the downscaling of the output.

    • bubo

      A silly problem, my bad,
      increasing number of iterations fixed my problem.
      Thanks anyway

  • Rui Guo