Abstract representations emerge naturally in neural networks trained to perform multiple tasks

• Published in: Nature communications Vol. 14; no. 1; pp. 1040 - 18
• Main Authors: Johnston, W. Jeffrey, Fusi, Stefano
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 23.02.2023; Nature Publishing Group; Nature Portfolio
• Subjects: 631/378/116/2395; 631/378/116/2396; Animals; Artificial neural networks; Brain - physiology; Cognitive ability; Humanities and Social Sciences; Humans; Machine learning; multidisciplinary; Neural networks; Neural Networks, Computer; Representations; Science; Science (multidisciplinary)
• ISSN: 2041-1723; 2041-1723
• DOI: 10.1038/s41467-023-36583-0

Humans and other animals demonstrate a remarkable ability to generalize knowledge across distinct contexts and objects during natural behavior. We posit that this ability to generalize arises from a specific representational geometry, that we call abstract and that is referred to as disentangled in machine learning. These abstract representations have been observed in recent neurophysiological studies. However, it is unknown how they emerge. Here, using feedforward neural networks, we demonstrate that the learning of multiple tasks causes abstract representations to emerge, using both supervised and reinforcement learning. We show that these abstract representations enable few-sample learning and reliable generalization on novel tasks. We conclude that abstract representations of sensory and cognitive variables may emerge from the multiple behaviors that animals exhibit in the natural world, and, as a consequence, could be pervasive in high-level brain regions. We also make several specific predictions about which variables will be represented abstractly. How animals learn to generalize from one context to another remains unresolved. Here, the authors show that the abstract representations that are thought to underlie this form of generalization emerge naturally in neural networks trained to perform multiple tasks.