Videos
See AI2's full collection of videos on our YouTube channel.Viewing 1-10 of 224 videos
Structure Modeling in Language Models
June 6, 2023 | Yuntian DengAbstract: We are approaching a future where text generation technologies will allow us to generate texts that are not only fluent at a surface level, but also coherent in their overall structure. To enable this future, my research focuses on evaluating and improving structure modeling in language models. In the…
When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories
June 6, 2023 | Alex MallenPresentation of ACL 2023 main conference long paper "When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories". Alex Mallen*, Akari Asai*, Victor Zhong, Rajarshi Das, Daniel Khashabi, Hannaneh Hajishirzi Despite their impressive performance on diverse tasks, large…
Enhancing the Reliability and Continual Improvement of Neural Dialogue Systems
June 5, 2023 | Prakhar GuptaNeural dialogue systems have made impressive advancements in natural language understanding and generation, yet challenges in reliability and continual improvement persist. In this talk, I will present my work focused on improving the reliability of dialogue systems while emphasizing continual enhancement and…
Toward Intelligent Writing Support Beyond Completing Sentences
May 23, 2023 | Chieh-Yang HuangAbstract: Writing is a challenging and demanding task that plays a critical role across various domains. Creative writing calls for innovative and attractive ideas, while scientific writing emphasizes clarity and a comprehensive presentation of concepts. Different fields present unique writing challenges and thus…
Towards robust long-form text generation systems
May 23, 2023 | Kalpesh KrishnaAbstract: Long-form text generation is an emerging AI technology that has seen extensive recent interest due to the success of ChatGPT. However, several key challenges need to be addressed before long-form generation systems can be practically deployed at scale. In this talk, I will present my research that…
Evaluating ethical and social risks from large models
April 25, 2023 | Laura Weidinger, Research Scientist, Deep MindAbstract: Large generative models create the potential for many beneficial use cases, but they also raise significant ethical and social risks of harm. How can we know whether a large model is aligned with our expectations? In this talk, we first present a taxonomy of ethical and social risks developed by a…
"Open" AI: considering the ethical upsides and downsides of "Open" AI development
April 6, 2023 | David Gray WidderAbstract: In this talk, I will discuss the ethical upsides and downsides of releasing AI openly. I will first present our FAccT’22 paper [1], where we interview contributors to an open source Deepfake tool about their sense of responsibility and agency to prevent harm. We show that open source licenses and…
The BigScience Workshop
March 30, 2023 | Yacine JerniteAbstract: The presentation will focus on presenting the evolution of the BigScience Workshop, its outcomes, and its focus on governance of the project itself and of its data and models. Bio: Yacine Jernite currently leads the ML and Society team at Hugging Face, where they work on ML systems governance at the…
Moving Forward by Moving Backward: Embedding Action Impact over Action Semantics
March 23, 2023 | Kuo-Hao Zeng, Luca Weihs, Roozbeh Mottaghi, Ali FarhadiThis video describes the Action Adaptive Policy by Zeng et al. (ICLR 2023). In the video, we discuss the formulation of motion drifts in point navigation and object navigation tasks. Additionally, we introduce the Action Adaptive Policy (AAP) to address these motion drifts. Finally, we demonstrate that AAP…
Recycling finetuned models to pretrain: on loss spaces, fusing and evolving pretraining
March 7, 2023 | Leshem ChoshenThis work will discuss the recent advancement in recycling finetuned models – Harnessing the data and computation invested in one or more models to collaboratively improve the pre-trained model they originated form. The work will also touch on our initial understandings of how and why fusing several models by…