Trustworthy machine learning by efficiently verifying compressed models
Abstract/Contents
- Abstract
- Autonomous decision-making systems are becoming ever more pervasive, and we are increasingly relying on these systems to perform actions for us. Previously, we mostly used algorithms for simple predictive tasks. Currently, we encounter them navigating sequential decision-making scenarios where they are crafted for choosing the sequence of actions that lead to, ideally, the maximal expected performance. With the widespread availability of data, increase in computational power, and advances in learning algorithms, machine learning is becoming a viable alternative to traditionally expert-crafted solutions. Machines are able to learn from data and build representations of the world that guide their actions. Recently, artificial neural networks have become very popular function approximators. Many astounding achievements of computer intelligence, from automated language translation to self-driving cars, are based on neural networks. In particular, their combination with Reinforcement Learning (RL) has enabled machines to learn the solution to complex sequential problems. Unlike traditional software, it is nearly impossible for a human to understand the logic that neural networks implements, making them opaque models and potentially precluding their use in safety or mission-critical applications. In many settings, running simulations is not enough to build confidence in them because one single failure can lead to catastrophic consequences. The work in this thesis addresses the challenge of building trust in machine-learned systems with neural network components. We begin by introducing neural network verification, a procedure that certifies whether a network upholds a desired property or not. We present recent advances to neural network verification, including our own contributions, and show that, despite progress, verification continues to be a very challenging problem and current algorithms struggle to scale to large networks. We then present an alternative approach that incorporates the need to verify into the design of the model. Simpler models are easier to verify and we show that some problems can be solved with binarized neural networks (BNNs), significantly simpler models with parameters that can be represented with 1 bit, with similar performance to their full precision counter-parts. We propose and demonstrate a simple mixed-integer programming approach to verify them and show that this approach scales well. Finally, we present a deep reinforcement learning algorithm, similar to Deep Q-Learning that uses BNNs as function approximators and we show that, again, this approach is able to tradeoff a small amount of performance and gain scalable verification.
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 | 2022; ©2022 |
Publication date | 2022; 2022 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | Lazarus Garcia, Christopher |
---|---|
Degree supervisor | Kochenderfer, Mykel J, 1980- |
Thesis advisor | Kochenderfer, Mykel J, 1980- |
Thesis advisor | Boyd, Stephen P |
Thesis advisor | Pilanci, Mert |
Degree committee member | Boyd, Stephen P |
Degree committee member | Pilanci, Mert |
Associated with | Stanford University, Institute for Computational and Mathematical Engineering |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Christopher Lazarus Garcia. |
---|---|
Note | Submitted to the Institute for Computational and Mathematical Engineering. |
Thesis | Thesis Ph.D. Stanford University 2022. |
Location | https://purl.stanford.edu/rs658gj7336 |
Access conditions
- Copyright
- © 2022 by Christopher Lazarus Garcia
- 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...