A computational framework for learning and transforming task representations
Abstract/Contents
- Abstract
- Human cognition is fundamentally flexible — we can adapt to novel tasks rapidly. We can sometimes adapt to a novel task without any direct experience on that task, based on its relationship to previous tasks. By contrast, while deep-learning models can achieve superhuman performance on many tasks, they are often unable to adapt to even slight task alterations. This ostensible inflexibility has led to criticism of deep learning models by cognitive scientists. I begin this dissertation by reviewing the literature on cognitive flexibility, and recent advances in building more flexible artificial intelligence systems. I provide a synthesis of these literatures, and outline the challenges that I believe remain. In particular, I focus on the ability to adapt to new tasks zero-shot — that is, without any data — based on their relationship to prior tasks. To address this challenge, I propose a general computational framework for adaptation to novel tasks based on their relationship to prior tasks. The framework is based on meta-mappings, higher-order tasks that transform basic tasks. I propose a parsimonious implementation of this framework in the form of homoiconic meta-mapping architectures. I demonstrate this framework across a wide variety of tasks and computational paradigms, ranging from regression to image classification and reinforcement learning. I compare to both human adaptability, and language-based approaches to zero-shot task performance. I show that meta-mapping is quite succesful, often achieveing 80-90% performance on a novel task, even when the new task directly contradicts prior experience. I further show that using this adaptation as a starting point can dramatically accelerate later learning on a task, and reduce the errors made on the way to mastery by nearly an order of magnitude. Thus, I suggest that meta-mapping can provide a computational basis for adapting to new tasks, and a starting point for efficient learning. This dissertation therefore provides a framework for building better cognitive models and more flexible artificial intelligence systems. In the final chapter, I review the broader contributions of this work to an ongoing discussion about the computational principles necessary for intelligence, and highlight possible future directions ranging from understanding mathematical cognition to neuroscience
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2020; ©2020 |
Publication date | 2020; 2020 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Lampinen, Andrew Kyle | |
---|---|---|
Degree supervisor | McClelland, James L | |
Thesis advisor | McClelland, James L | |
Thesis advisor | Ganguli, Surya, 1977- | |
Thesis advisor | Goodman, Noah | |
Degree committee member | Ganguli, Surya, 1977- | |
Degree committee member | Goodman, Noah | |
Associated with | Stanford University, Department of Psychology. |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Andrew Kyle Lampinen |
---|---|
Note | Submitted to the Department of Psychology |
Thesis | Thesis Ph.D. Stanford University 2020 |
Location | electronic resource |
Access conditions
- Copyright
- © 2020 by Andrew Kyle Lampinen
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...