Global SummitAI EventsEngage
Neural Network

Unlearning Toxicity in Multimodal Foundation Models & Learning to design protein-protein interactions with enhanced generalization

OnlineDiscovery - From Molecules to Models

  • Register

    Watch

    * Register (or log in) to the Neural Network to add this session to your agenda or watch the replay

    • Date
    3 February 2025
    Timeframe
    16:00 - 17:30
    Duration
    90 minutes
    Share this session

    Part 1 (Rita Cucchiara): Foundation Models, pretrained on extremely large unknown source of data, contain in their embed space many information representation that are not desirable, since are inappropriate, privacy unrespectful, un-ethic, violent or unsafe for many possible users (e.g. talking about nudity), and toxic in large sense.  Some techniques of filtering in input or before output can mitigate this effect, although filters can be removed, especially in open-source models.  This talk discusses some direction for removing the learned knowledge in embedded spec by unlearning by redirection, i.e. suggesting a redirection of some points in the embedded space with a specific finetuning. We explore solutions in the challenging case of multimodal embedded space, created by contrastive learning such as in CLIP for connecting text and image data, for removing some toxic concepts that can be used for both retrieval and generative downstream tasks with a prototype called SAFE-CLIP, recently presented at ECCV2024.  The results, part of the EU ELIAS project, could represent a first exploratory direction to provide next generation of responsible Multimodal LLMs and foundational models, avoiding the use or the abuse of Less-Suitable-For- Work (LSFW) concepts in image and text generation, and thus supporting future generation of sustainable AI.

    Part 2 (Josef Sivic): Discovering mutations enhancing protein-protein interactions (PPIs) is critical for advancing biomedical research and developing improved therapeutics. While machine learning approaches have substantially advanced the field, they often struggle to generalize beyond training data in practical scenarios. The contributions of this work are three-fold. First, we construct PPIRef, the largest and non-redundant dataset of 3D protein-protein interactions, enabling effective large-scale learning. Second, we leverage the PPIRef dataset to pre-train PPIformer, a new SE(3)-equivariant model generalizing across diverse protein-binder variants. We fine-tune PPIformer to predict effects of mutations on protein-protein interactions via a thermodynamically motivated adjustment of the pre-training loss function. Finally, we demonstrate the enhanced generalization of our new PPIformer approach by outperforming other state-of-the-art methods on new, non-leaking splits of standard labeled PPI mutational data and independent case studies optimizing a human antibody against SARS-CoV-2 and increasing the thrombolytic activity of staphylokinase. 

    Share this session
    • Days
      Hours
      Min
      Sec
    Add to Calendar
    In partnership with
    Related stories

    Harnessing AI to Connect Every School Worldwide

    By Celia Pizzuto Connecting every school on the planet to the internet by 2030 is an ambitious goal, but with...
    Global Summit

    Leveraging AI for Humanitarian and Peacekeeping Efforts

    By Celia Pizzuto As technological advancements accelerate, the challenge remains to ensure these innovations reach those in need. At the...
    Global Summit

    Tristan Harris at the AI for Good Global Summit: The...

    By Rukevwe Omoro, edited by Celia Pizzuto In today's fast-paced world of artificial intelligence (AI), the conversation about its growth...
    Global Summit

    Are you sure you want to remove this speaker?