I was reading Hinton & Salakhutdinov's 2006 paper on autoencoders and Figure 3b (as well as S4 in the Supplementary Materials) are rather confusing due the absence of any axis? Anyone knows how the figures are obtained?
Intuitively I figured that the axis are the probabilities [0,1] of the 2 hidden units on the top-most layer and the recovered digit is based on a downpass using the associated probabilities?
[link][3 comments]