Instantiating human inductive biases in machines via metalearningONR White Paper Tracking Number: 23-000004312
Abstract
Despite significant advances in machine learning and artificial intelligence research in the last decade, humans remain the best example we have of systems that are able to extract generalizable abstract knowledge from very small amounts of data. The key to this ability is having strong inductive biases # those factors other than the data that influence learning. Human learning can be captured by probabilistic models of cognition, in which inductive biases are expressed in terms of probability distributions over hypotheses that can include complex abstractions such as grammars, logical formulas, and programs. However, there is a substantial differencebetween the scale of cognitive science experiments and machine learning applications, so models of human cognition that are effective in the lab # for example, models based on computationally costly methods such as Bayesian inference and program induction # do not naturally translate into systems that scale to machine learning applications. This proposal describes a research program intended to overcome this challenge, using metalearning as a formal framework for bridging probabilistic models of cognition and machine learning methods based on deep neural networks. Metalearning is a class of techniques in which a system is trained to perform a set of related tasks. While each task is performed separately, the system has hyperparameters that are shared across tasks, such as the initial weights used in artificial neural networks, and a metalearning algorithm updates those parameters with the goal of improving theperformance across all of the tasks. Typically, metalearning is applied to standard tasks from machine learning (e.g. image classification) and focuses on developing models that can extract the common structure from a set of tasks, resulting in a system that has inductive biases aligned with those tasks. The proposed research expands upon this use of metalearning by sampling tasks from the prior distributions ofprobabilistic models of cognition that are calibrated against human behavior. This provides a way to distill human inductive biases into machine learning systems that can operate at scale and be applied to realistic data. This approach will be applied to problems that pose a challenge to contemporary machine learning systems # learning logical and geometric concepts from small amounts of data and learning when the number of classes is unknown. The research will also explore two other ways of using metalearning to improve performance in these settings: training systems to perform efficient probabilistic inference in nonparametric Bayesian models, and use of auxiliary tasks to create appropriate inductive biases even in settings where probabilistic models of humancognition not available. The results will expand the capabilities of machine learning systems, produce systems that are better aligned with their human operators, and provide new tools for predicting human behavior.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 29, 2023
- Source ID
- N000142312510
Entities
People
- Thomas L. Griffiths
Organizations
- Office of Naval Research
- Trustees of Princeton University
- United States Navy