Frederick L. Hovde Distinguished Lecturer: Wojciech Szpankowski

"Foundations of Structural, Temporal, and Semantic Information"

April 8, 2026 - 3:00 PM Eastern Time

Location: DSAI 1069

Lecture: 3 p.m.

Reception: 4 p.m.

Abstract

Shannon’s information theory has provided the intellectual basis for communication and storage systems for more than five decades. While extraordinarily successful, it was intentionally designed to ignore several aspects of information that are now central to modern data and AI systems: complex structures (such as those in networks and geometry), temporal dynamics, and meaning or semantics—famously set aside by Shannon as being outside the scope of engineering. Today, information plays a far broader role than simply enabling reliable transmission and storage. It underlies how we analyze networks (e.g., what are the emergent properties of agent-based networks, how systems biology models code phenotypes, how social networks propagate information), learn from data (e.g., markers and progression of disease from observational data, inferring material properties from structure databases, etc.), and reason about knowledge itself (without hallucination). This shift calls for a richer theory of information that explicitly incorporates structure, time, semantics, and reasoning.

In this talk, I describe recent progress toward such a theory. I begin by showing that temporal information in evolving networks can be recovered solely from the network’s final structure, even when temporal annotations are absent. I then show how to establish fundamental limits on the amount of information contained in common data structures, with particular emphasis on networks. Finally, I turn to semantic information and propose a novel framework that connects information theory with logic. This is of particular importance, given our focus on the ability (or lack thereof) of AI models to reason and make logical inferences. The key idea is that if one statement can be logically deduced from another, then—given sufficient computational power— it carries no new information. From this perspective, understanding semantic information requires identifying when different statements are meaningfully equivalent. We introduce a new mathematical framework built on this principle and show how it naturally extends Shannon’s classical communication model by equipping the receiver with the ability to perform logical inference. This theory moves beyond raw data transmission to enable reasoning about data itself, reframing communication as the transfer of knowledge-bearing information and providing a principled foundation for formalizing knowledge transfer and learning.