Class-Specifed Denoising Auto Encoder
Denoising Auto Encoder (DAE) is an artificial neural network trained to reconstruct a data sample from a corrupted version of that sample. This thesis presents Class-Specific DAE (csDAE), a modular method for training a DAE. In the csDAE method, we train several small DAEs on disjoint subgroups of the dataset and then combine them into one DAE. The motivation behind this approach is that DAE implicitly estimates the data-generating density; therefore it may be easier to estimate a distribution of several smaller and maybe simpler sets than of a large and complicated one. We compared csDAE performance on a multiclass classification task with standard DAE. Our results show that csDAE and standard DAE perform equally well. We show that csDAE significantly improves training time compared to to standard DAE.
* This research was carried out towards the M.Sc. Degree in Electrical Engineering at Bar-Ilan University, under the supervision of Prof. Gal Chechik