Canberra Deep Learning Meetup

In recent months, the release of Open AI's O1 preview has sparked increased debate around the term "reasoning" as applied to LLMs. This talk will illustrate the monumental strides that have been made in mathematical reasoning by AI at a technical level in the past year.

In a paper published in January in Nature, DeepMind introduce AlphaGeometry, an AI system that solves complex geometry problems at a level approaching a human Olympiad gold-medalist.

In July, DeepMind announced a generalised method for solving maths Olympiad problems called AlphaProof, including an improved AlphaGeometry 2.

The talk will discuss the methodology of the January paper in detail, with the goal of understanding how AlphaGeometry could reach the ability of the world's best Olympiad contestants without human demonstration.

Finally we will discuss the implications for the future of AI reasoning in more general domains, both within and beyond mathematics.

Useful link:

AlphaGeometry: https://www.nature.com/articles/s41586-023-06747-5

Join us for an exciting meetup event as we explore the groundbreaking research presented in the paper "Accurate Structure Prediction of Biomolecular Interactions with AlphaFold 3." This year's Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper for their pioneering work on AlphaFold 2, which transformed our ability to predict complex protein structures.

In this session, we'll delve into AlphaFold 3 and its predecessor to understand how its advanced deep-learning architecture has achieved unprecedented accuracy in predicting the structures of protein complexes, nucleic acids, and small molecules. These advancements mark significant progress in therapeutic design and molecular biology. Don’t miss this opportunity to engage with the latest in scientific innovation and technology!

Paper link: https://www.nature.com/articles/s41586-024-07487-w

This latest paper from Deepmind explored the potential of LLM on self-correction. They propose a method, SCoRe, a multi-turn online reinforcement learning (RL) approach designed to enhance the self-correction ability of large language models (LLMs) using self-generated data. Traditional methods for training self-correction often require multiple models or external supervision and have shown limited effectiveness. SCoRe improves on this by addressing the limitations of supervised fine-tuning (SFT), which often leads to distribution mismatches or ineffective correction behaviors at test time.

SCoRe trains the model under its own distribution of correction traces, applying regularization to guide the model toward an effective self-correction strategy. The approach involves two key phases: an initial RL phase for generating a stable policy and the use of reward bonuses to boost self-correction during training. When applied to Gemini 1.0 Pro and 1.5 Flash models, SCoRe significantly improves self-correction performance, achieving state-of-the-art results with a 15.6% improvement on the MATH benchmark and a 9.1% improvement on the HumanEval benchmark.

paper: https://arxiv.org/abs/2409.12917

Join us at the Canberra Deep Learning Meetup to delve into the transformation of LLM models into Mixture of Experts. This paper discussed the methodology of improving the dense models into sparse Mixture of Experts. This yields improvement in performance without much extra computational requirement.

Paper: https://arxiv.org/pdf/2410.07524