PUBLIC MARKS from ogrisel with tags layer & ai

A long-term goal of Machine Learning research is to solve highly complex "intelligent" tasks, such as visual perception auditory perception, and language understanding. To reach that goal, the ML community must solve two problems: the Deep Learning Problem, and the Partition Function Problem. There is considerable theoretical and empirical evidence that complex tasks, such as invariant object recognition in vision, require "deep" architectures, composed of multiple layers of trainable non-linear modules. The Deep Learning Problem is related to the difficulty of training such deep architectures. Several methods have recently been proposed to train (or pre-train) deep architectures in an unsupervised fashion. Each layer of the deep architecture is composed of an encoder which computes a feature vector from the input, and a decoder which reconstructs the input from the features. A large number of such layers can be stacked and trained sequentially, thereby learning a deep hierarchy of features with increasing levels of abstraction. The training of each layer can be seen as shaping an energy landscape with low valleys around the training samples and high plateaus everywhere else. Forming these high plateaus constitute the so-called Partition Function problem. A particular class of methods for deep energy-based unsupervised learning will be described that solves the Partition Function problem by imposing sparsity constraints on the features. The method can learn multiple levels of sparse and overcomplete representations of data. When applied to natural image patches, the method produces hierarchies of filters similar to those found in the mammalian visual cortex. An application to category-level object recognition with invariance to pose and illumination will be described (with a live demo). Another application to vision-based navigation for off-road mobile robots will be described (with videos). The system autonomously learns to discriminate obstacles from traversable areas at long range.

In the 1980's, new learning algorithms for neural networks promised to solve difficult classification tasks, like speech or object recognition, by learning many layers of non-linear features. The results were disappointing for two reasons: There was never enough labeled data to learn millions of complicated features and the learning was much too slow in deep neural networks with many layers of features. These problems can now be overcome by learning one layer of features at a time and by changing the goal of learning. Instead of trying to predict the labels, the learning algorithm tries to create a generative model that produces data which looks just like the unlabeled training data. These new neural networks outperform other machine learning methods when labeled data is scarce but unlabeled data is plentiful. An application to very fast document retrieval will be described.