Research Projects

Transfer Learning for Image Classification via Joint Sparse Approximation

*** THIS PROJECT IS NO LONGER ACTIVE ***

Trevor Darrell, Ariadna Quattoni¹, Michael Collins and Xavier Carreras

An ideal image classifier should be able to exploit complex high dimensional feature representations even when only a few labeled examples are available for training. To achieve this goal we developed a transfer learning model that can leverage data from related categories to make learning easier from small samples. Our model is based on the observation that related categories might be learnable using only a small subset of shared relevant features. To find these features we propose to train the categories jointly with a shared regularization penalty that minimizes the total number of features involved in the approximation.

The resulting optimization problem can be formulated as a linear program and solved with an out of the shelve package. While this approach is feasible for small problems, it becomes untractable for large datasets; the second part of our work addresses this problem by developing an optimization algorithm for joint sparse approximation whose complexity is linear with the number of training examples and O(n log n) with n being the number of parameters of the joint model. We use this model on a news-topic prediction task where the goal is to predict whether an image belongs to a particular news topic.

[1]

Ariadna Quattoni, Michael Collins, Trevor Darrell, Xavier Carreras, Transfer Learning for Image Classification via joint sparse approximation, CVPR 2008

¹MIT