en
0.25
0.5
0.75
1.25
1.5
1.75
2
Higher Order Contractive auto-encoder
Published on Nov 30, 20113822 Views
We propose a novel regularizer when training an autoencoder for unsupervised feature extraction. We explicitly encourage the latent representation to contract the input space by regularizing the norm
Related categories
Chapter list
Higher Order Contractive Auto-Encoder (CAE+H)00:00
Overview of the presentation00:12
The basics00:42
Reconstruction error01:30
First Order Contractive Auto-Encoder01:52
Higher Order CAE02:43
Why penalize the derivative's norm?04:20
Geometric interpretation of the CAE+H05:14
Local space contraction - 105:46
Local space contraction - 206:17
Contraction Ratio - 107:16
Contraction Ratio - 208:24
Local contraction09:18
Reconstruction vs. Contraction penalty09:37
Approximating the manifold using the encoder's mapping11:10
Manifold learning context12:24
Local charts of the manifold12:58
Local coordinates and saturation14:03
Formal de nition of the atlas14:55
Visualizing the tangents15:30
Overcomplete representation16:54
CAE+H features17:18
CAE+H: What happens when we go deep?18:02
CIFAR10 performance18:16
Future Work18:27
Thanks to ...19:11