Seminario Machine Learning
Better vision modeling, and better artificial neural networks, by considering dendritic processing
Ponente: Marcelo Bertalmío (Instituto de Óptica, CSIC)Fecha: viernes 17 de abril de 2026 - 12:00Lugar: Aula Roja, IFT
Resumen:
Vision models are now developed mostly by training artificial neural networks (ANNs), with whom they share the same structure, motivated by the neuroscience of the mid-20th century. However, ANNs have critical limitations that biological systems do not have, and we hypothesize that a key factor in these limitations comes from the overly simplistic neuron model that ANNs have been using since their inception, which assumes that dendrites behave linearly. We have developed a more realistic, biologically plausible neuron model that takes into account dendritic processing, and we show how this model has a number of significant advantages both for vision modeling and for ANNs in general, allowing to build systems that generalize better, are more robust and need less parameters and less data to determine them.
