iiinova — where your digital self keeps evolving

Mahalakshmi Sabanayagam

@maha-saba.bsky.social

PhD student @ TU Munich. Interested in theory of deep learning and graph based learning. web: mahalakshmi-sabanayagam.github.io

Posts Replies Media

Mahalakshmi Sabanayagam @maha-saba.bsky.social · 24/04/2025

To hear all about the work @iclr-conf.bsky.social, Poster: 25.04 from 3:00 to 5:30 pm Talk: 27.04 at 3:30 pm in the VerifAI workshop verifai-workshop.github.io/schedule/ Paper: arxiv.org/abs/2412.00537 Looking forward to meeting and discussing with fellow researchers in Singapore! 🙂 [6/6]

021

Mahalakshmi Sabanayagam @maha-saba.bsky.social · 24/04/2025

⏩ Our experimental results on (graph) neural networks in semi-supervised node classification uncover a novel phenomenon: a plateauing of robustness for intermediate perturbations (for more details and other interesting results, see our paper arxiv.org/abs/2412.00537). [5/6]

110

Mahalakshmi Sabanayagam @maha-saba.bsky.social · 24/04/2025

🥳 The resulting certificate in both cases is a Mixed Integer Linear Program (MILP) and scales with the number of labelled samples. Thus, making it very much suitable for semi-supervised learning, which naturally has a low labeling rate. [4/6]

110

Mahalakshmi Sabanayagam @maha-saba.bsky.social · 24/04/2025

🔥 How do we handle it? We leverage the Neural Tangent Kernel as it describes the training dynamics of wide NNs under certain conditions, and develop exact certification strategies for both sample-wise and collective certification. [3/6]

100

Mahalakshmi Sabanayagam @maha-saba.bsky.social · 24/04/2025

🤔 All exact certificates for neural nets target test-time robustness. Here, we ask what if the training data, e.g. labels are corrupted. 🤯 Challenge: This brings additional complexity of handling the training dynamics of NNs during the certification of predictions. [2/6]

100

Mahalakshmi Sabanayagam @maha-saba.bsky.social · 24/04/2025

🎉Excited to announce our #ICLR2025 Spotlight! 🚀 @lukasgosch.bsky.social and I will be presenting our paper on the first exact certificate against label poisoning for neural nets and graph neural nets. Joint work with Stephan Guennemann and Debarghya Ghoshdastidar. 👇[1/6]

110