We introduce a new framework for learning in severely resource-constrained settings. Our technique delicately amalgamates the representational richness of multiple linear predictors with the sparsity of Boolean relaxations, and thereby yields classifiers that are compact, interpretable, and accurate. We provide a rigorous formalism of the learning problem, and establish fast convergence of the ensuing algorithm via relaxation to a minimax saddle point objective.

Summary of paper: The paper addresses the problem of learning in resource constrained setting using multiple linear predictors (max of linear predictors like DNF). The problem of learning k-DNFs (a subset of multiple linear predictord) is known to be hard however to solve this non-convex problem, they use various relaxations to formulate this as a minmax saddle point objective which they solve using standard mirror-prox algorithm. The paper then goes on to show that the approach is fruitful in practice through experiments. Comments for authors: In the introduction (line 28-31), you seem to sugest that you can solve the problem of k-DNFs however you do not actually give any guarantees for the final learnt hypothesis that your algorithm outputs, please make that clear. Requiring a different weight vector for each positive example, this seems very costly and potentially avoidable.

We introduce a new framework for learning in severely resource-constrained settings. Our technique delicately amalgamates the representational richness of multiple linear predictors with the sparsity of Boolean relaxations, and thereby yields classifiers that are compact, interpretable, and accurate. We provide a rigorous formalism of the learning problem, and establish fast convergence of the ensuing algorithm via relaxation to a minimax saddle point objective. Papers published at the Neural Information Processing Systems Conference.