BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Vincent Szolnoky (Chalmers University of Technology & University o f Gothenburg) DTSTART:20240306T121500Z DTEND:20240306T130000Z DTSTAMP:20260422T155153Z UID:gbgstats/45 DESCRIPTION:Title: On the Interpretability of Regularisation for Neural Networks Through Mo del Gradient Similarity\nby Vincent Szolnoky (Chalmers University of T echnology & University of Gothenburg) as part of Gothenburg statistics sem inar\n\nLecture held in MVL14.\n\nAbstract\nMost complex machine learning and modelling techniques are prone to over-fitting and may subsequently ge neralise poorly to future data. Artificial neural networks are no differen t in this regard and\, despite having a level of implicit regularisation w hen trained with gradient descent\, often require the aid of explicit regu larisers. We introduce a new framework\, Model Gradient Similarity (MGS)\, that (1) serves as a metric of regularisation\, which can be used to moni tor neural network training\, (2) adds insight into how explicit regularis ers\, while derived from widely different principles\, operate via the sam e mechanism underneath by increasing MGS\, and (3) provides the basis for a new regularisation scheme which exhibits excellent performance\, especia lly in challenging settings such as high levels of label noise or limited sample sizes.\n LOCATION:https://researchseminars.org/talk/gbgstats/45/ END:VEVENT END:VCALENDAR