at NeurIPS 2020

During this live Q&A, Ben will be discussing his and OpenAI’s major contributions in this paper, as well as where we fell short. His work was mainly on training data, eval memorization, and the eval suite. He will offer deep dives on these sections.

Prafulla has worked on both Language Models are Few-Shot Learners and Jukebox: A Generative Model for Music focusing on generative models, in particular on scaling them to high dimensional data like audio and images. His most recent work is Jukebox, which showcases the ability of neural nets to produce music with singing.

We demonstrate that scaling up language models greatly improves task-agnostic, few-shot performance, sometimes even becoming competitive with prior state-of-the-art fine-tuning approaches. Specifically, we train GPT-3, an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model, and test its performance in the few-shot setting.

Andrew will be demonstrating the capabilities of the API via the interactive playground and tools for Semantic Search.

Chat with members of our Open-Endedness team about anything related to their work and research at virtual Table #6 at the NeurIPS Social.

As language models become more powerful, training and evaluation are increasingly bottlenecked by the data and metrics used for a particular task. In this work, we show that it is possible to significantly improve summary quality by training a model to optimize for human preferences.

Reinforcement learning agents typically fall into uncooperative equilibria when trained in social dilemma environments. We explore whether randomized and uncertain social preferences can pressure agents into more cooperative equilibria.

If you missed his poster session last night, come chat with Bowen at our booth!

Ben will describe OpenAI’s major contributions in this paper, as well as where we fell short. His work was mainly on training data, eval memorization, and the eval suite. He will offer deep dives on these sections.

Andrew will be demonstrating the capabilities of the API via the interactive playground and tools for Semantic Search.

This workshop aims to bring together researchers from all the different communities and topics that fall under the umbrella of meta-learning. We expect that the presence of these different communities will result in a fruitful exchange of ideas and stimulate an open discussion about the current challenges in meta-learning as well as possible solutions.

The exponential growth of AI research has led to several papers floating on arxiv, making it difficult to review existing literature. Despite the huge demand, the proportion of survey & analyses papers published is very low due to reasons like lack of a venue and incentives. Our Workshop, ML-RSA provides a platform and incentivizes writing such types of papers. It meets the need of taking a step back, looking at the sub-field as a whole and evaluating actual progress.

Lilian has been working on teaching robots to solve a wide variety of tasks via RL training in simulation and sim2real transfer to the physical world. Her recent work uses asymmetric self-play to train a single, goal-conditioned policy that can solve many robotic manipulation tasks, including tasks with previously unseen goals and objects. Come and chat with Lilian about the paper "Asymmetric self-play for automatic goal discovery in robotic manipulation.

Join OpenAI’s Miles Brundage—with Bryan McCann, Colin Raffel, Natalie Schulter, Zeerak Waseem, Rosie Campbell—as they discuss responsible publication following growing concerns with both harmful research impact and research conduct in computer science and research published at NeurIPS.

The research on this paper grew out of the work Kamal did with OpenAI’s Scholars program, mentored by Natasha Jaques. In this paper (winner of a best paper award at the Cooperative AI Workshop), they show that independent RL agents can learn social policies that allow them to learn from experts in their environment. The learned social learning behavior allows them to accomplish tasks that would otherwise be prohibitively difficult, and to perform well in unseen social tasks.