Attention: A machine learning perspective

Abstract

We review a statistical machine learning model of top-down task driven attention based on the notion of `gist'. In this framework we consider the task to be represented as a classification problem with two sets of features - a gist of coarse grained global features and a larger set of low-level local features. Attention is modeled as the choice process over the low-level features given the gist. The model takes its departure in a classical information theoretic framework for experimental design. This approach requires the evaluation over marginalized and conditional distributions. By implementing the classifier within a Gaussian Discrete mixture it is straightforward to marginalize and condition, hence, we obtained a relatively simple expression for the feature dependent information gain - the top-down saliency. As the top-down attention mechanism is modeled as a simple classification problem, we can evaluate the strategy simply by estimating error rates on a test data set. We illustrate the attention mechanism on a simple simulated visual domain in which the choice is over nine patches in which a binary pattern has to be classified. The performance of the classifier equipped with the attention mechanism is almost as good as one that has access to all low-level features and clearly improving over a simple `random attention' alternative.