BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Noemi Montobbio (IIT) DTSTART:20230109T131500Z DTEND:20230109T141500Z DTSTAMP:20260423T010636Z UID:MaML/6 DESCRIPTION:Title: Eme rgence of Lie Symmetries in Functional Architectures Learned by CNNs\n by Noemi Montobbio (IIT) as part of Mathematics and Machine Learning\n\n\n Abstract\nConvolutional Neural Networks (CNNs) are a powerful tool providi ng outstanding performances on image classification tasks\, based on an ar chitecture designed in analogy with information processing in biological v isual systems. The functional architectures of the early visual pathways h ave often been described in terms of geometric invariances\, and several s tudies have leveraged this framework to investigate the analogies between CNN models and biological mechanisms. Remarkably\, upon learning on natura l images\, the translation-invariant filters of the first layer of a CNN h ave been shown to develop as approximate Gabor functions\, resembling the orientation-selective receptive profiles found in the primary visual corte x (V1). With a similar approach\, we modified a standard CNN architecture to insert computational blocks compatible with specific biological process ing stages\, and studied the spontaneous development of approximate geomet ric invariances after training the network on natural images. In particula r\, inserting a pre-filtering step mimicking the Lateral Geniculate Nucleu s (LGN) led to the emergence of a radially symmetric profile well approxim ated by a Laplacian of Gaussian\, which is a well-known model of receptive profiles of LGN cells. Moreover\, we introduced a lateral connectivity ke rnel acting on the feature space of the first network layer. We then studi ed the learned connectivity as a function of relative tuning of first-laye r filters\, thus re-mapping it into the roto-translation space. This analy sis revealed orientation-specific patterns\, which we compared qualitative ly and quantitatively with established group-based models of V1 horizontal connectivity.\n LOCATION:https://researchseminars.org/talk/MaML/6/ END:VEVENT END:VCALENDAR