University of Minnesota
A Statistical Perspective on Machine Intelligence: Lifelong Learning and Reliable Reasoning
Abstract
dc:description.abstractThis thesis studies two capabilities central to machine intelligence, lifelong learning and reliable causal reasoning, from a statistical perspective. Lifelong learning requires a system to learn a sequence of tasks while improving over time and retaining what it has already learned, yet learning new tasks typically degrades performance on earlier ones. We study it from two sides. First, we formulate lifelong learning as nonparametric estimation under recurrence and derive its minimax rate, which separates the cost of estimating each recurring distribution from the cost of identifying which one generates the current data. Second, we propose DriftNet, a biologically inspired method that induces representational drift, a phenomenon observed in biological learning, to store and retrieve task-specific knowledge without task identities. DriftNet outperforms existing lifelong-learning methods on image-classification and natural-language tasks and scales to language models with billions of parameters. The second capability, reliable causal reasoning, concerns the validity of conclusions rather than the retention of knowledge. Large language models reach confident but unreliable conclusions about cause and effect. We introduce CausalPitfalls, a benchmark spanning six categories of causal error that measures whether a model's conclusions are stable under reformulation and whether they distinguish real effects from spurious ones. Reliability is low under direct prompting, and code-assisted prompting benefits only the strongest models.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Du, Jin
Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Citation
- Du, Jin. (2026). A Statistical Perspective on Machine Intelligence: Lifelong Learning and Reliable Reasoning. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/282296.