Talk by Kirill Neklyudov (University of Montreal)

The talk presents a novel and elegant mathematical framework for combining probabilistic models via set operations. The speaker’s derivations are rigorous, carefully handling singularities through limits and renormalization. The motivation is clear: enabling compositional generation without retraining, which is highly relevant for practical applications like drug design. The approach is theoretically sound, building on measure theory and probability. However, the talk is primarily a conceptual presentation; empirical validation is not shown in the transcript. The speaker acknowledges that union requires more complex operations, but does not elaborate. The connection to diffusion models is briefly mentioned but not detailed, assuming audience familiarity. The sources are not explicitly cited in the transcript, but the talk is part of a Simons Institute workshop, lending credibility. Overall, the talk offers a valuable contribution to the field, though further empirical evidence would strengthen its impact. The title is generic but accurate. The audience question about the complement definition is well-addressed, clarifying the practical choice for generative modeling.

The title is generic but accurately reflects the content: a technical talk by Kirill Neklyudov at the Simons Institute.

7/10

The talk presents a novel mathematical framework for probabilistic set operations applied to generative models, with rigorous derivations and clear explanations. The speaker is a researcher at Mila, a reputable institution. However, the talk is a presentation of ongoing work without peer-reviewed publication details, and the transcript is incomplete, limiting full verification.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction and overview of the talk on steering generative models.
• 02:00Definition of the full set as a Boltzmann distribution over all possible states.
• 05:00Introduction of probabilistic set definition using indicator functions.
• 10:00Derivation of intersection operation using square roots and algebraic manipulation.
• 15:00Definition of complement via limit and renormalization, addressing singularities.
• 20:00Derivation of set difference and its practical importance for safety in drug design.
• 25:00Discussion of training data collection and conditional distributions.
• 30:00Connection to diffusion models and flow matching, with visual illustration.
• 35:00Application to protein generation, combining models conditioned on different properties.
• 40:00Conclusion and potential future directions.

• Simons Institute talk page — Official page for the talk, providing context and possibly slides.

• Simons Institute talk page — Official page for the talk, providing context and possibly slides.