Modern deep learning is exceptionally good at seeing patterns, but it often sees too much. As models scale, they increasingly begin to see “ghosts", which are nuisance factors that haunt the data and masquerade as true signals. These ghosts appear as stereotypes in social data, as overwhelming thermal emission in physical sensor data, or as rigid concept associations in generative models. All of these artifacts obscure the true causal structure of the world.
In this talk, I argue that scaling alone cannot exorcise these ghosts. Instead, we must learn the art of unseeing them. I will present a unified framework for Invariant Representation Learning that formalizes this unseeing as "structural surgery" on the underlying causal graph. This approach mathematically severs the dependence on nuisance factors while preserving the true signal.
I will demonstrate how this single theoretical approach addresses three seemingly distinct challenges:
1. Exorcising Social Ghosts: How to surgically erase sensitive attributes from latent representations to ensure fairness and controllability, without destroying model utility.
2. Exorcising Physical Ghosts: How thermodynamic laws can be used as causal priors to disentangle direct heat emission from surface texture in thermal imaging, enabling AI to see clearly through darkness.
3. Exorcising Conceptual Ghosts: How decomposing score functions in diffusion models allows us to "unsee" spurious co-occurrences, unlocking compositional generalization for novel creation.
By moving from passive correlation mining to active structural enforcement, this work lays the foundation for AI systems that are not fooled by the ghosts in the data, but are robust, fair, creative, and grounded in the causal structure of the real world.
Vishnu Naresh Boddeti is an Associate Professor in the Department of Computer Science and Engineering at Michigan State University. His research focuses on building AI systems with provable guarantees, moving beyond "black box" scaling to models that are fair, private, and physically grounded.
His work spans three interconnected pillars: (1) Responsible AI, where he develops invariant representation learning methods to audit and mitigate bias in foundation models; (2) Physics-Informed AI, which integrates physical laws with AI models; and (3) Secure AI, designing AI systems that operate on homomorphically encrypted user data for real-world deployment. His research has been featured on the cover of Nature and recognized with multiple awards, including the 2024 IEEE-CCF Best Paper Award and the 2023 IEEE-TBIOM Best Student Paper Award. He currently serves as a Senior Area Editor for IEEE TIFS.