Introspective Intelligence: Computing, Acting and Discovering Under Uncertainty

Abstract: As artificial intelligence (AI) systems move toward long-horizon operation in physical and scientific environments, they require reliable mechanisms for recognizing when their predictions, internal representations or actions are becoming unreliable. This talk examines uncertainty as a computational capability that spans the computing stack.

Modern AI models exhibit measurable signatures associated with uncertainty in their outputs, internal activations, temporal consistency and execution trajectories, enabling uncertainty to be estimated in some cases from a single forward pass rather than through costly ensemble or repeated sampling. Making such estimation practical also requires efficient physical implementation.

We demonstrate compute-in-memory circuits that integrate Monte Carlo sampling with memory operations as well as programmable heterojunction transistors whose Gaussian-like transfer characteristics can directly implement functions useful for probabilistic inference, motivating materials and device primitives for native belief computation. We then consider uncertainty not simply as an estimated quantity but as a signal for decision and resource control.

Calibrated uncertainty can determine when a system should abstain, verify, acquire additional observations or replan, while also dynamically allocating computation, precision, memory and sensing resources according to the difficulty of a decision. In adaptive visual tracking, for example, uncertainty-guided model selection reduced average computation by 58.2% while maintaining detection quality within 0.4%. For robots and long-horizon agents, trajectory-level uncertainty can provide early indications of failure, identify where execution begins to diverge from expected behavior and trigger targeted verification or recovery.

Finally, uncertainty provides a natural acquisition signal for discovery, guiding the selection of experiments, simulations, measurements and candidate designs that are expected to provide the greatest information gain, closing the loop from intelligent systems back to the materials and computing substrates on which they are built.

Bio: Amit Ranjan Trivedi is a professor of electrical and computer engineering, and of computer science by courtesy, at the University of Illinois Chicago, where he leads the AEON Lab. His honors include the National Science Foundation CAREER Award, the Institute of Electrical and Electronics Engineers (IEEE) Electron Devices Society Ph.D. Fellowship, an Intel Researcher of the Year Award as one of ten recipients in 2024 and a Best Paper Award at the IEEE International Conference on Artificial Intelligence Circuits and Systems. He received his Ph.D. in electrical and computer engineering from the Georgia Institute of Technology, where his doctoral dissertation received the Best Ph.D. Thesis Award, and previously worked at IBM’s Semiconductor Research and Development Center.