Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Rapid unsupervised alignment with the natural image manifold

Ananya Passi¹, Brian S Robinson², Michael Bonner¹; ¹Johns Hopkins University, ²Johns Hopkins University Applied Physics Laboratory

Presenter: Ananya Passi

There is a stark contrast between the nature of feature learning in biological and artificial vision. While brains learn without explicit supervision and with little data, deep neural networks require supervised feedback and massive training sets. Here we show that a surprisingly simple unsupervised learning algorithm can yield large improvements in the brain alignment of a deep vision model. Specifically, we trained a network in which each layer learns to compress its representations onto the principal modes of variance for natural images—a form of local learning that does not require backpropagation or supervision. Using a relatively small sample of training images, this unsupervised learning algorithm strongly improves the network’s ability to predict the image-evoked fMRI responses of visual cortex, and it makes downstream learning on an image-classification task more efficient. Remarkably, after an initial unsupervised-learning phase, the first half of the network’s layers can be frozen with little impact on the ability to learn image classification. Together, these findings suggest that a parsimonious learning algorithm—operating locally and without supervision—may be sufficient to induce the features of early-to-mid-level vision and may accelerate the learning of downstream task-specific functions.

Topic Area: Visual Processing & Computational Vision

Extended Abstract: Full Text PDF