Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Unsupervised learning of human-interpretable and task-relevant visual features from very little data

Ananya Passi¹, Brian S Robinson¹, Michael Bonner¹; ¹Johns Hopkins University

Presenter: Ananya Passi

Deep learning models typically require large amounts of labeled training data and do not explain how the human visual system can learn effectively from very little data. Fully unsupervised learning, considered important in biological vision, is not commonly used in current deep neural networks. We investigated whether integrating unsupervised efficient coding principles into neural networks allows learning in situations with limited data. We developed a deep hierarchical network in which each layer learns an efficient code directly from the activations of its inputs. This fully unsupervised approach does not involve any supervised or self-supervised task, nor does it require backpropagation. We used this model to address two main questions: (1) Can a hierarchy of efficient coding yield perceptually relevant visual features? (2) Can efficient coding support rapid downstream learning of new categories from limited data? Our findings show that after training on just thousands of images, the deep efficient coding network learns visual features that human observers can readily recognize. In contrast, observers had difficulty recognizing the features encoded in key comparison networks, including a supervised network trained on the same data and an untrained network. Furthermore, we found that when networks are trained to classify images in challenging low-data settings, their performance can be substantially boosted by an initial unsupervised phase using the efficient coding algorithm. Together, these findings reveal the remarkable effectiveness of efficient-coding principles for learning a hierarchy of behaviorally relevant features and enabling data-efficient visual learning.

Topic Area: Methods, Tools, Theory & Neural Coding