en-es
en
0.25
0.5
0.75
1.25
1.5
1.75
2
Learning Representations: A Challenge for Learning Theory
Published on 2013-08-0920598 Views
Perceptual tasks such as vision and audition require the construction of good features, or good internal representations of the input. Deep Learning designates a set of supervised and unsupervised met
Related categories
Presentation
Learning Representations: A Challenge For Learning Theory00:00
Learning Representations: a challenge for AI, ML, Neuroscience, Cognitive Science01:35
Architecture of “Mainstream” Machine Learning Systems02:44
This Basic Model has not evolved much since the 50's03:47
The Mammalian Visual Cortex is Hierarchical04:05
Let's be inspired by nature, but not too much05:20
Trainable Feature Hierarchies: End-to-end learning06:57
Do we really need deep architectures?07:31
Why would deep architectures be more efficient?08:47
Deep Learning: A Theoretician's Nightmare? - 109:41
Deep Learning: A Theoretician's Nightmare? - 211:34
Deep Learning: A Theoretician's Paradise?12:54
Deep Learning and Feature Learning Today13:56
In Many Fields, Feature Learning Has Caused a Revolution15:04
Convolutional Networks16:25
Early Hierarchical Feature Models for Vision16:55
The Convolutional Net Model18:16
Feature Transform - 118:20
Feature Transform - 219:03
Convolutional Network (ConvNet)19:12
Convolutional Network Architecture19:14
Convolutional Network (vintage 1990)19:16
“Mainstream” object recognition pipeline 2006-201221:33
Tasks for Which Deep Convolutional Nets are the Best21:48
Simple ConvNet Applications with State-of-the-Art Performance22:28
Prediction of Epilepsy Seizures from Intra-Cranial EEG22:33
Epilepsy Prediction22:36
ConvNet in Connectomics22:38
Object Recognition [Krizhevski, Sutskever, Hinton 2012] - 123:15
Object Recognition [Krizhevski, Sutskever, Hinton 2012] - 224:26
Another ImageNet-trained ConvNet [Zeiler & Fergus 2013]24:32
Object Recognition on-line demo25:29
ConvNet trained ImageNet [Zeiler & Fergus 2013]25:39
Features are generic: Caltech 25625:51
Features are generic: PASCAL VOC 201226:43
Labeling every pixel with the object it belongs to27:11
Scene Parsing/Labeling: ConvNet Architecture27:44
Scene Parsing/Labeling: Performance - 129:01
Scene Parsing/Labeling: Performance - 229:58
Scene Parsing/Labeling: SIFT Flow dataset - 130:27
Scene Parsing/Labeling: SIFT Flow dataset - 230:28
Scene Parsing/Labeling - 130:32
Scene Parsing/Labeling - 230:37
Scene Parsing/Labeling - 330:39
Scene Parsing/Labeling - 430:41
NYU RGB-Depth Indoor Scenes Dataset32:09
Scene Parsing/Labeling on RGB+Depth Images32:13
Semantic Segmentation on RGB+D Images and Videos32:21
Unsupervised Learning32:34
Discovering the Hidden Structure in High-Dimensional Data - 133:49
Discovering the Hidden Structure in High-Dimensional Data - 234:26
Basic Idea of Feature Learning36:03
Energy-Based Unsupervised Learning39:05
Energy Function foro a Manifold40:07
Sparse Coding & Sparse Modeling41:04
Strategies to Learn an Energy Function42:31
How to Speed Up Inference in a Generative Model?44:21
Idea: Train a “simple” function to approximate the solution44:34
Regularized Encoder-Decoder Model44:50
PSD: Basis Functions on MNIST45:01
Predictive Sparse Decomposition (PSD): Training45:03
Learned Features on natural patches45:27
Better Idea: Give the “right” structure to the encoder45:29
LISTA: Train We and S matrices46:18
Discriminative Recurrent Sparse Auto-Encoder46:36
DrSAE Discovers manifold structure of handwritten digits46:52
Convolutional Sparse Coding47:12
Convolutional PSD47:33
Convolutional Sparse Auto-Encoder on Natural Images48:06
Using PSD to Train a Hierarchy of Features - 148:11
Using PSD to Train a Hierarchy of Features - 248:36
Using PSD to Train a Hierarchy of Features - 348:46
Using PSD to Train a Hierarchy of Features - 448:51
Using PSD to Train a Hierarchy of Features - 548:54
Detection49:28
Pedastrian Detection49:57
Unsupervised pre-training with convolutional PSD 50:30
Detection Picture - 150:35
Detection Picture - 251:01
Unsupervised Learning: Invariant Features51:13
Learning Invariant Features with L2 Group Sparsity51:15
Groups are local in a 2D Topographic Map51:20
Topographic Maps51:46
Image-level training, local filters but no weight sharing51:50