Using Natural Language to Guide Meta-Learning Agents towards Human-like Inductive Biases

Sreejan Kumar, Ishita Dasgupta, Michael Hu, Raja Marjieh, Robert D. Hawkins, Nathaniel Daw, Jonathan Cohen, Karthik R Narasimhan, Thomas L. Griffiths

TL;DR: We show the conditions in which one can guide meta-reinforment learning agents to produce human-like behavior using natural language descriptions of hidden task structure.

Abstract: Inductive biases are a key component of human intelligence, allowing people to acquire, represent, and use abstract knowledge. Although meta-learning has emerged as an approach to endowing neural networks with inductive biases, agents trained via meta-learning can use very different strategies compared to humans. We show that co-training these agents on predicting human-generated natural language task descriptions guides them toward human-like inductive biases that more appropriately capture the structure of the task distribution as humans see it. We further show that the level of abstraction at which humans write these descriptions influences the size of the effect. This work provides a foundation for investigating how to collect task descriptions at the appropriate level of abstraction to leverage for approximating human-like learning of structured representations in neural networks.

Track: Non-Archival (will not appear in proceedings)