Learning Feature Hierarchies for Object Recognition
Candidate: Koray Kavukcuoglu
Advisor: Yann LeCun

Abstract

In this thesis we study unsupervised learning algorithms for training feature extractors and building deep learning models. We propose sparse-modeling algo- rithms as the foundation for unsupervised feature extraction systems. To reduce the cost of the inference process required to obtain the optimal sparse code, we model a feed-forward function that is trained to predict this optimal sparse code. Using an efficient predictor function enables the use of sparse coding in hierarchical models for object recognition. We demonstrate the performance of the developed system on several recognition tasks, including object recognition, handwritten digit classification and pedestrian detection. Robustness to noise or small variations in the input is a very desirable property for a feature extraction algorithm. In order to train locally-invariant feature extractors in an unsupervised manner, we use group sparsity criteria that promote similarity between the dictionary elements within a group. This model produces locally-invariant representations under small pertur- bations of the input, thus improving the robustness of the features. Many sparse modeling algorithms are trained on small image patches that are the same size as the dictionary elements. This forces the system to learn multiple shifted versions of each dictionary element. However, when used convolutionally over large im- ages to extract features, these models produce very redundant representations. To avoid this problem, we propose convolutional sparse coding algorithms that yield a richer set of dictionary elements, reduce the redundancy of the representation and improve recognition performance.