GPT-3 and Three Collaborative Studies Win NeurIPS 2020 Best Paper Awards
OpenAI's massive GPT-3 language model shares the prestigious NeurIPS 2020 Best Paper Award with research from Politecnico di Milano, Carnegie Mellon University, and UC Berkeley. The winners highlight major advancements in few-shot learning, no-regret learning dynamics, and data summarization.
OpenAI’s groundbreaking GPT-3 language model paper shares the NeurIPS 2020 Best Paper Award with a no-regret learning dynamics study from Politecnico di Milano and Carnegie Mellon University, along with a data summarization project from UC Berkeley. The NeurIPS organizing committee announces these winners to kick off the thirty-fourth Conference on Neural Information Processing Systems, which draws over 18,000 participants to a virtual gathering.
The GPT-3 paper demonstrates that scaling up language models to 175 billion parameters greatly improves task-agnostic, few-shot performance, making it competitive with prior state-of-the-art fine-tuning approaches. This massive autoregressive model contains ten times more parameters than any previous non-sparse language model and achieves remarkable results without requiring any gradient updates or fine-tuning for specific tasks.
To accommodate the massive global audience, NeurIPS 2020 organizers design a highly accessible virtual event schedule featuring two six-hour daily sessions that start at 5am PT and 5pm PT. Paper authors choose their preferred session to align with their local time zones, and attendees select their preferred bandwidth to ensure smooth participation regardless of their internet connection speeds.