Visual Perception with Deep Learning

Google Tech TalksApril, 9 2008ABSTRACTA long-term goal of Machine Learning research is to solve highycomplex "intelligent" tasks, such as visual perception auditoryperception, and language understanding. To reach that goal, the MLcommunity must solve two problems: the Deep Learning Problem, and thePartition Function Problem.There is considerable theoretical and empirical evidence that complextasks, such as invariant object recognition in vision, require "deep"architectures, composed of multiple layers of trainable non-linearmodules. The Deep Learning Problem is related to the difficulty oftraining such deep architectures.Several methods have recently been proposed to train (or pre-train)deep architectures in an unsupervised fashion. Each layer of the deeparchitecture is composed of an encoder which computes a feature vectorfrom the input, and a decoder which reconstructs the input from thefeatures. A large number of such layers can be stacked and trainedsequentially, thereby learning a deep hierarchy of features withincreasing levels of abstraction. The training of each layer can beseen as shaping an energy landscape with low valleys around thetraining samples and high plateaus everywhere else. Forming thesehigh plateaus constitute the so-called Partition Function problem.A particular class of methods for deep energy-based unsupervisedlearning will be described that solves the Partition Function problemby imposing sparsity constraints on the features. The method can learnmultiple levels of sparse and overcomplete representations ofdata. When applied to natural image patches, the method produceshierarchies of filters similar to those found in the mammalian visualcortex.An application to category-level object recognition with invariance topose and illumination will be described (with a live demo). Anotherapplication to vision-based navigation for off-road mobile robots willbe described (with videos). The system autonomously learns todiscriminate obstacles from traversable areas at long range.This is joint work with Y-Lan Boureau, Sumit Chopra, Raia Hadsell,Fu-Jie Huang, Koray Kavakcuoglu, and Marc'Aurelio Ranzato.Speaker: Yann Le CunComputational and Biological Learning Lab, Courant Institute of Mathematical Sciences,New York University.

Channel: People & Blogs
Uploaded: December 31, 1969 at 3:59 pm
Author: googletechtalks

Length: 57:25
Rating: 4.69
Views: 7721

Tags: education  engedu  google  googletechtalks  talk  talks  techtalk  techtalks  

Video Comments