Microsoft is a proud sponsor of the 38th Conference on Neural Information Processing Systems (NeurIPS) 2024, a premier international event for the exchange of ideas in the field of machine learning and artificial intelligence. This year’s conference, which brings together researchers, industry leaders, and practitioners, will showcase Microsoft’s contributions to the advancement of AI across diverse domains.
At NeurIPS 2024, Microsoft will have a significant presence, with over 100 accepted papers authored by Microsoft researchers and their collaborators. These contributions include five oral presentations and 19 spotlight sessions. The research spans a broad range of topics, with a central theme: improving the efficiency, scalability, and robustness of machine learning models while addressing real-world challenges.
Lidong Zhou, Managing Director of Microsoft Research Asia, will be a keynote speaker at the conference. Attendees are encouraged to visit Microsoft’s booth (#445) to learn more about these advancements and engage with researchers.
Key research areas highlighted by Microsoft researchers include:
- Efficiency and Scalability: Research focused on optimizing machine learning models for improved performance and resource utilization.
- Robustness: Studies aimed at making AI systems more reliable and less susceptible to errors.
- Human-Centric Interaction: Investigations into how AI can better understand and interact with people.
- Cultural Considerations: Research that addresses how to account for cultural diversity and biases in AI models.
Spotlight Presentations and Selected Papers:
The following papers represent a selection of the research being presented by Microsoft:
- Not All Tokens Are What You Need for Pretraining: This project introduces Rho-1, a novel language model that employs selective language modeling. Unlike conventional models that predict every subsequent token, Rho-1 concentrates on training with tokens aligned with the desired distribution. This model earned a “Best Paper Runner Up Award.”
- Reinforcement Learning Under Latent Dynamics: Toward Statistical and Algorithmic Modularity: A project looking at reinforcement learning with general latent dynamics and develops efficient reductions to adapt latent Markov decision process (MDP) algorithms for complex observations.
- CVQA: Culturally-diverse Multilingual Visual Question Answering Benchmark: This research introduces a benchmark that involves native speakers and cultural experts in the data collection process. CVQA includes culturally driven images and questions from 30 countries across four continents, covering 31 languages and 13 scripts.
- VASA-1: Lifelike Audio-Driven Talking Faces Generated in Real Time: This project presents VASA-1, a framework for generating lifelike talking faces that synchronize lip movements with speech and capture facial nuances and natural head motions.
- You Only Cache Once: Decoder-Decoder Architectures for Language Models: This paper details a decoder-decoder architecture for LLMs that reduces GPU memory usage by caching key-value pairs only once, while retaining global attention.
Microsoft’s participation also extends to the Machine Learning for Health (ML4H) Symposium, which is co-located with NeurIPS 2024. Microsoft researchers will present four papers at ML4H, demonstrating the company’s commitment to advancing AI applications in healthcare to improve medical imaging and clinical workflows. Further details on Microsoft’s presence at NeurIPS 2024 and related resources are readily available.
