BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:Asia/Kolkata BEGIN:STANDARD DTSTART:19451014T230000 TZOFFSETFROM:+0630 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20260204T182946Z UID:936D876B-7E10-4824-B96A-946C69360DB9 DTSTART;TZID=Asia/Kolkata:20240611T143000 DTEND;TZID=Asia/Kolkata:20240611T153000 DESCRIPTION:The success of deep learning (DL) and convolutional neural netw orks (CNN) has also highlighted that NN-based analysis of signals and imag es of large sizes poses a considerable challenge\, as the number of NN wei ghts increases exponentially with data volume – the so called Curse of D imensionality. In addition\, the largely ad-hoc fashion of their developme nt\, albeit one reason for their rapid success\, has also brought to light the intrinsic limitations of CNNs - in particular those related to their black box nature. To this end\, we revisit the operation of CNNs from firs t principles and show that their key component – the convolutional layer – effectively performs matched filtering of its inputs with a set of te mplates (filters\, kernels) of interest. This serves as a vehicle to estab lish a compact matched filtering perspective of the whole convolution-acti vation-pooling chain\, which allows for a theoretically well founded and p hysically meaningful insight into the overall operation of CNNs. This is s hown to help mitigate their interpretability and explainability issues\, t ogether with providing intuition for further developments and novel physic ally meaningful ways of their initialisation. Such an approach is next ext ended to Graph CNNs (GCNNs)\, which benefit from the universal function ap proximation property of NNs\, pattern matching inherent to CNNs\, and the ability of graphs to operate on nonlinear domains. GCNNs are revisited sta rting from the notion of a system on a graph\, which serves to establish a matched-filtering interpretation of the whole convolution-activation-pool ing chain within GCNNs\, while inheriting the rigour and intuition from si gnal detection theory. This both sheds new light onto the otherwise black box approach to GCNNs and provides well-motivated and physically meaningfu l interpretation at every step of the operation and adaptation of GCNNs. I t is our hope that the incorporation of domain knowledge\, which is centra l to this approach\, will help demystify CNNs and GCNNs\, together with es tablishing a common language between the diverse communities working on De ep Learning and opening novel avenues for their further development.\n\nSp eaker(s): Dr. Danilo P. Mandic\, \n\nVirtual: https://events.vtools.ieee.o rg/m/423821 LOCATION:Virtual: https://events.vtools.ieee.org/m/423821 ORGANIZER:ieee.sps.sb.iitkgp@gmail.com SEQUENCE:16 SUMMARY:IEEE SPS SBC Webinar: Interpretable Convolutional NNs and Graph CNN s (By Dr. Danilo P. Mandic) URL;VALUE=URI:https://events.vtools.ieee.org/m/423821 X-ALT-DESC:Description: <br /><p><strong id="docs-internal-guid-cfd97124-7f ff-1532-aec4-9521a0a1b96d">&nbsp\;</strong>The success of deep learning (D L) and convolutional neural networks (CNN) has also highlighted that NN-ba sed analysis of signals and images of large sizes poses a considerable cha llenge\, as the number of NN weights increases exponentially with data vol ume &ndash\; the so called Curse of Dimensionality. In addition\, the larg ely ad-hoc fashion of their development\, albeit one reason for their rapi d success\, has also brought to light the intrinsic limitations of CNNs - in particular those related to their black box nature. To this end\, we re visit the operation of CNNs from first principles and show that their key component &ndash\; the convolutional layer &ndash\; effectively performs m atched filtering of its inputs with a set of templates (filters\, kernels) of interest. This serves as a vehicle to establish a compact matched filt ering perspective of the whole convolution-activation-pooling chain\, whic h allows for a theoretically well founded and physically meaningful insigh t into the overall operation of CNNs. This is shown to help mitigate their interpretability and explainability issues\, together with providing intu ition for further developments and novel physically meaningful ways of the ir initialisation. Such an approach is next extended to Graph CNNs (GCNNs) \, which benefit from the universal function approximation property of NNs \, pattern matching inherent to CNNs\, and the ability of graphs to operat e on nonlinear domains. GCNNs are revisited starting from the notion of a system on a graph\, which serves to establish a matched-filtering interpre tation of the whole convolution-activation-pooling chain within GCNNs\, wh ile inheriting the rigour and intuition from signal detection theory. This both sheds new light onto the otherwise black box approach to GCNNs and p rovides well-motivated and physically meaningful interpretation at every s tep of the operation and adaptation of GCNNs. It is our hope that the inco rporation of domain knowledge\, which is central to this approach\, will h elp demystify CNNs and GCNNs\, together with establishing a common languag e between the diverse communities working on Deep Learning and opening nov el avenues for their further development.</p> END:VEVENT END:VCALENDAR