Papers in NeurIPS 2020

We are very happy to announce that the members of our team have had two accepted papers in NeurIPS 2020!

More particularly:

  • Amina Mollaysa is the first author of "Goal-directed Generation of Discrete Structures with Conditional Generative Models" (pre-print available here),  a work done with Brooks Paige, and Alexandros Kalousis. The work proposes an approach to conditional generation of discrete structures, such as chemical molecules, which instead of sampling from a learned model such that a reward measure is maximized, samples from a reward based distribution and maximizes the likelihood of high reward. Such an approach avoids the limitations of reinforce like algorithms, which use score-function based estimators that exhibit high variance.
  • Jason Ramapuram, in the context of his internship at DeepMind, worked on "Self-Supervised MultiModal Versatile Networks" (pre-print available here) which introduces MMV FAC. MMV FAC is a large scale versatile self-supervised multi-modal model that can be used for audio, video, text and image downstream tasks. MMV FAC is trained with 16 years of Youtube data with ASR annotations and presents state-of-the-art results on numerous small scale datasets such as UCF101 (video), HMDB51 (video) and ESC-50(audio) in comparison to previous self-supervised work.