Niket Tandon from the Allen Institute for AI presented a talk at MBZUAI on enabling large language models to focus on human needs and continuously learn from interactions. He proposed a memory architecture inspired by the theory of recursive reminding to guide models in avoiding past errors. The talk addressed who to ask, what to ask, when to ask and how to apply the obtained guidance. Why it matters: The research explores how to align LLMs with human feedback, a key challenge for practical and ethical AI deployment.
MBZUAI researchers identified 'self-referencing causal cycles' in LLM training data that can mitigate the 'reversal curse,' where LLMs struggle with information presented in reverse order. The study, to be presented at ACL, explains that the transformer architecture's unidirectional token generation causes this issue. By leveraging the repetitive nature of information in training texts, the team developed an efficient solution to improve LLM performance. Why it matters: Overcoming the reversal curse can significantly enhance LLM accuracy and reliability, especially in tasks requiring bidirectional reasoning and understanding of context.
MBZUAI hosted a talk by Visiting Associate Professor Adrian Bors on continuously streaming AI and the challenge of catastrophic forgetting. The talk covered approaches to continual learning like expanding mixtures of models and generative replay mechanisms. Results were presented on image classification and generation tasks. Why it matters: Continual learning is crucial for AI systems to adapt to new environments and real-world data without forgetting previous knowledge.
Researchers at MBZUAI have developed DynaMMo, a dynamic model merging method for efficient class incremental learning using medical images. DynaMMo merges multiple networks at different training stages using lightweight learnable modules, reducing computational overhead. Evaluated on three datasets, DynaMMo achieved a 10-fold reduction in GFLOPS compared to existing dynamic methods with a 2.76 average accuracy drop.
Liangming Pan from UCSB presented research on building reliable generative AI agents by integrating symbolic representations with LLMs. The neuro-symbolic strategy combines the flexibility of language models with precise knowledge representation and verifiable reasoning. The work covers Logic-LM, ProgramFC, and learning from automated feedback, aiming to address LLM limitations in complex reasoning tasks. Why it matters: Improving the reliability of LLMs is crucial for high-stakes applications in finance, medicine, and law within the region and globally.
Pascal Fua from EPFL presented an approach to implementing convolutional neural nets that output complex 3D surface meshes. The method overcomes limitations in converting implicit representations to explicit surface representations. Applications include single view reconstruction, physically-driven shape optimization, and bio-medical image segmentation. Why it matters: This research advances geometric deep learning by enabling end-to-end trainable models for 3D surface mesh generation, with potential impact on various applications in computer vision and biomedical imaging in the region.
Keith Ross, Dean of Computer Science, Data Science and Engineering at NYU Shanghai, will be giving a talk on recent advances in Deep Reinforcement Learning (DRL). The talk will review DRL breakthroughs and discuss algorithmic research on DRL for high-dimensional state and action spaces, with applications to robotic locomotion. Ross's research interests include deep reinforcement learning, Internet privacy, peer-to-peer networking, and computer network modeling. Why it matters: Reinforcement learning is a core area of AI research in the GCC region, and a talk by a prominent researcher can help inform and inspire local researchers.