#AAAI2024 workshops round-up 3: human-centric representation learning, and AI to accelerate science and engineering

In this third round-up article of the workshops at AAAI 2024, we hear from the organisers of the workshops on:

• Human-Centric Representation Learning
• AI to Accelerate Science and Engineering

Human-Centric Representation Learning
Organisers: Dimitris Spathis, Aaqib Saeed, Ali Etemad, Stefanos Laskaridis, Shohreh Deldari, Sana Tonekaboni, Patrick Schwab, Shyam Tailor, Ian Tang.

The Human-Centric Representation Learning workshop at AAAI 2024 brought together researchers who are broadly interested in modern representation learning for human-centric data. Representation learning has become a key research area in machine learning and artificial intelligence, with the goal of automatically learning useful representations of data for a wide range of tasks. Powerful models like GPT4 or Stable Diffusion are trained in a self-supervised way in order to learn generalized representations. However, traditional representation learning approaches often fail to consider the human perspective and context, leading to representations that may not be interpretable or meaningful to humans. Here are some highlights from the workshop:

• Accepted papers spanned a diverse range of topics in cutting edge AI research and applications. This included computer vision, multimodal learning, fairness and ethics considerations, interpretability and explainability of models, learning effective representations, continual learning, generative modeling techniques, and novel applications in healthcare among others.
• We gave awards to three papers which share the common goal of aligning AI models, especially large language models, with human values, preferences and social intelligence. One proposes techniques for improved controllability of language model outputs through activation steering, allowing humans to guide model behavior. Another explores hybrid natural language and feedback signals to fine-tune models towards satisfying human feedback during training itself. The third takes a fundamental perspective, studying how AI can learn representations mirroring human conceptualization and reasoning, positing this as a path to imbue AI with robust human values and social skills. Collectively, these works recognize the critical need for advanced AI capabilities to be directed responsibly, cohering with human preferences, value systems and societal norms.
• We had five keynote talks from academic and industry experts who covered a wide range of AI research topics. Ishan Misra discussed using powerful diffusion models as multimodal world models that can generate videos and integrate with large language models. Jimeng Sun presented work on leveraging generative AI for improving various aspects of clinical trial development. Marinka Zitnik covered research towards developing universal foundation models for time series data across diverse domains. Neil Zeghidour introduced audio language models that unify audio analysis and synthesis using neural codecs and autoregressive sequence modeling. Ahmad Beirami talked about alignment techniques to fine-tune large language models to satisfy human preferences while minimally perturbing the base model. Despite their distinct focus areas, the keynotes collectively highlighted the frontiers of generative AI capabilities, multimodal learning, domain-specific adaptations of foundation models, human-centered controls and applications across areas like healthcare and audio processing. They exemplified the rapid progress towards more capable, general and human-compatible AI systems that can synthesize rich data modalities while adhering to human-specified constraints and real-world requirements.

By Dimitris Spathis and Aaqib Saeed

AI to Accelerate Science and Engineering
Organisers: Aryan Deshwal, Jana Doppa, Syrine Belakaria, Kaiyan Qiu, Kaiyan Qiu.

The 3rd Annual AAAI Workshop on AI to Accelerate Science and Engineering (AI2ASE) brought together researchers from AI and diverse science/engineering communities to achieve the following goals: 1) Identify and understand the challenges in applying AI to specific science and engineering problems, 2) Develop, adapt, and refine AI tools for novel problem settings and challenges, 3) Community-building and education to encourage collaboration between AI researchers and domain area experts.

• The workshop had 34 papers accepted from widely different scientific and engineering domains including materials science, astrophysics, power engineering, biodiversity sciences. Interestingly, a large part of them also focused on machine learning driven decision-making aspect rather than just solving standard static prediction tasks.
• The workshop has a main theme each year. This year the theme was AI for Materials and Manufacturing which is very critical area of research for United States. There were 6 invited speakers with expertise ranging from core materials science and manufacturing to artificial intelligence.
• The main themes of speakers’ talks and panel discussion centered around three key challenges and ideas: how to efficiently acquire information for goal-directed search under resource constraints (adaptive experimental design/Bayesian optimization/entropic sampling), how to incorporate physical constraints for reasoning with deep models, how to explain AI based decisions for domain scientists in order to discover new scientific knowledge (via better visualization techniques for instances)?

By Aryan Deshwal

tags: AAAI, AAAI2024