Zoom Meeting: https://umass-amherst.zoom.us/j/5464223779
Abstract:
As increasingly sensitive decision making problems become automated using models trained by machine learning algorithms, it is becoming important to design training algorithms that provide assurance that the models they produce will be well behaved. However, the task of designing such algorithms is challenging, and there are several obstacles that can prevent existing safe machine learning strategies from being useful in practice. For example, because of the many types of adverse effects that can arise when deploying a model, such as potential for physical harm, unfair bias, and more, it is important for safe training algorithms to be flexible in terms of how "bad behavior" is quantified. Unfortunately, many existing approaches make limiting assumptions about how safety should be quantified, preventing them from being useful in many problem settings. Also, in many cases, models are trained using data that is not truly representative of what will be encountered once the model is deployed. Consequently, the assurances or guarantees provided by many existing safe algorithms are not helpful, since they may fail to hold once the model is deployed. In this proposal defense, I will consider these challenges, and address the design of practical machine learning algorithms that provide high-confidence guarantees on the behavior of the trained models.
First, I will consider the problem of providing high-confidence guarantees based on general definitions of safety, and propose algorithms that allow the user to provide these definitions as simple text input to training algorithms. Next, I will address the challenge of establishing high-confidence guarantees that hold even when the data used to train the model is not representative of the environment the model will be deployed into. Because of the many ways in which the training and deployment environments can differ, we identify two distinct settings, and propose algorithms for addressing each.
Advisor: Phil Thomas