ProbXiv
Explorersign in
Problem archiveProblem record

Statement

In particular, there is some duality in the way the minimization over f \in F is performed, as one can decide to either (i) simultaneously minimize over fenc and fdec , which is the standard way to proceed when training neural networks; (ii) minimize first over the encoder part, which allows to reformulate the min- imization as the well-known problem of finding principal manifolds; (iii) minimize first over the decoder part, which is natural when thinking of the reconstruction error as some total variance to be decomposed using a conditioning on the values of the encoder. 1.2. Reformulating autoencoders with conditional expectations. We dis- cuss here how to reformulate the training of autoencoders with conditional expec- tations, and provide alternative interpretations to the reconstruction error. We minimize the reconstruction error (3) by first minimizing over the decoder part for a given encoding function, as already considered in [7]. This approach is natural in molecular dynamics, as it is reminiscent of free energy computations [3, 11] where average quantities are computed for a fixed value of the collective variable fenc . From a mathematical viewpoint, it corresponds to introducing conditional averages associated with fixed values of the encoder. The loss function for unsupervised least-squares can be rewritten as h i h i 2 2 inf E kX − f (X)k = inf inf E kX − fdec ◦ fenc (X)k f \inF fenc \inFenc fdec \inFdec h i 2 (4) = inf E X − gf⋆enc ◦ fenc (X) , fenc \inFenc

Record

Comments

Nothing has been published against this problem yet, and nothing has been said about it.

Sign in with an institutional address to take part in the discussion. Reading every thread stays open to everyone.

Sign in

Solve with an agent

Open the statement in a chat, with the problem and the ground rules already written into the prompt.

This opens a third-party site. Nothing is posted back to ProbXiv and nothing you write there is recorded here — what a model gives you is an attempt, which a person still has to check.