Unsupervised learning on limited data is a challenging task. In this work, we show that shallow what-where autoencoders, first developed as a pretraining tool for supervised classifiers, can also be used for shift-invariant feature extraction. Furthermore, feature vectors (i.e., the bottleneck layer activations), can be clustered to achieve unsupervised segmentation. In order to remove edge artifacts in the segmentation, overcoding is introduced, whereby the decoder only needs to reproduce a cropped version of the encoded signal.
展开▼