Generative models for classical and quantum matter
Location: CECAM-HQ-EPFL, Lausanne, Switzerland
Organisers
Due to the high number of applications compared to the number of spots available, we have anticipated the registration deadline to October 14th at midnight CEST.
Description
Generative modeling, a broad term that encompasses many machine learning techniques to generate random variables that resemble samples from a given target distribution, are being heavily explored in the computational molecular sciences. For example, pioneering works using these techniques in a chemical context include efforts to sample thermal distributions of molecular systems [1], enhance free energy methods [2], evaluate intractable integrals arising in quantum field theories [3], carrying out direct simulations, among many other applications. The main paradigms that have been exploited in the molecular sciences include likelihood estimation with normalizing flows (i.e., training from knowledge of the energy alone) or training from an existing dataset, typically collected using molecular dynamics simulations [4]. At present, it is not clear whether substantive accelerations relative to explicit simulations can be obtained for every system [5].
Indeed, while there is palpable excitement about generative modeling techniques, a number of determinative questions remain:
- How do we build generative models for chemical systems that generalize beyond a training dataset?
- Which generative models are most appropriate for a given task?
- How do we train transferable models that are appropriate for a large class of systems?
- Which criteria or metrics should we use to evaluate the quality of generative models?
This workshop seeks to highlight and explore these questions by bringing together a group of computational scientists with diverse expertise. In particular, we hope to integrate discussion of neural network architectures that are well-suited to chemical data [6,7] and discuss the implications of such architectures for generalization of generative models. The workshop, ultimately, will serve to disseminate recent advances across communities and reflect on the next steps to tackle several critical challenges in the field.
References
Marylou Gabrié (École Polytechnique) - Organiser
Switzerland
Giuseppe Carleo (EPFL) - Organiser
United States
Grant Rotskoff (Stanford University) - Organiser