Search Papers | Poster Sessions | All Posters
Poster C100 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
A local unsupervised learning algorithm for building a visual hierarchy
Ananya Passi1, Michael Bonner1 (); 1Johns Hopkins University
Deep neural networks (DNNs) are the leading computational models of visual cortex, but they are trained using a biologically implausible mechanism that backpropagates a learning signal through the entire network hierarchy. We developed an approach for building a hierarchy of visual features using only local unsupervised learning, without the need for backpropagation. In our algorithm, each layer of a DNN contains a bottleneck in which representations are compressed and then expanded again. Learning is fully unsupervised, with each layer learning only to compress its inputs. This parsimonious algorithm yields representations that are competitive with conventional DNNs at predicting visual cortex representations up to intermediate layers. This work identifies a new approach for learning a visual hierarchy that is consistent with principles of learning in biology, requires no image labels or tasks, and may be sufficient to account for large fraction of visual cortex representations.
Keywords: convolutional neural networks encoding models deep learning visual cortex