A Neural Mechanism for Transfer Learning

Abstract

Transfer learning, the application of knowledge from a training domain to a novel domain, is a critical advantage for human intelligence and an eagerly sought goal for artificial intelligence. Here, we describe a neural mechanism for transfer learning based on compositional representation. In monkeys trained on a small alphabet of letter-like stimuli, neurons in high-level visual cortex showed increased selectivity specifically for compositional letter elements (limbs, junctions, terminations) and their interactions (combinations of limbs, etc.) and not for whole letters. Strikingly, the enhanced selectivity for the familiar compositional elements and their interactions transferred to novel letters constructed from the same elements, making the novel letters more discriminable and providing information about their structure. Compositionality is a powerful strategy for knowledge transfer in domains like vision where objects often share common elements. Our results reveal a brain mechanism for visual transfer learning that could be adapted for artificial vision.