AI has entered a pivotal point of remarkable efficacy across domains, yet the fundamental principles explaining how it works remain open questions.
AI models can be seen as a series of embeddings that sequentially transform the data to uncover and exploit its intrinsic structure. We aim to understand how AI discovers and leverages such structure.
Such understanding would bring transformative benefits: it would allow us to design more reliable and interpretable AI systems, extend their use to safety-critical fields such as healthcare and climate science, and accelerate scientific discovery by revealing the hidden structures of complex data.
Publications |
---|
K. Rupashree, S. Baskar, and D. Pimentel-Alarcón. "Some neural networks inherently preserve subspace clustering structure". Under review. 2025. |
S. Baskar, K. Rupashree, and D. Pimentel-Alarcón. "Deep union completion". AAAI Conference on Artificial Intelligence. 2025. [Link] |
K. Rupashree, S. Baskar, and D. Pimentel-Alarcón. "Latent union completion". IEEE International Symposium on Information Theory (ISIT). 2025. [Link] |