BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Ismael Lemhadri (Stanford University) DTSTART:20200709T150000Z DTEND:20200709T160000Z DTSTAMP:20260423T003255Z UID:ProbStat/7 DESCRIPTION:Title: LassoNet: A Neural Network with Feature Sparsity\nby Ismael Lemhadri (Stanford University) as part of Probability & Statistics (IST-CEMAT\, FC -CEAUL\, ULisbon)\n\n\nAbstract\nMuch work has been done recently to make neural networks more interpretable\, and one obvious approach is to arrang e for the network to use only a subset of the available features. In linea r models\, Lasso (or $\\ell_1$-regularized) regression assigns zero weight s to the most irrelevant or redundant features\, and is widely used in dat a science. However the Lasso only applies to linear models. Here we introd uce LassoNet\, a neural network framework with global feature selection. O ur approach enforces a hierarchy: specifically a feature can participate i n a hidden unit only if its linear representative is active. Unlike other approaches to feature selection for neural nets\, our method uses a modifi ed objective function with constraints\, and so integrates feature selecti on with the parameter learning directly. As a result\, it delivers an enti re regularization path of solutions with a range of feature sparsity. On s ystematic experiments\, LassoNet significantly outperforms state-of-the-ar t methods for feature selection and regression. The LassoNet method uses p rojected proximal gradient descent\, and generalizes directly to deep netw orks. It can be implemented by adding just a few lines of code to a standa rd neural network.\n LOCATION:https://researchseminars.org/talk/ProbStat/7/ END:VEVENT END:VCALENDAR