5 papers
ReSAE: Residualized Sparse Autoencoders for Multi-Layer Transformer Interventions
Prathyush Poduval, Calvin Yeung, Neel Desai +1
Sparse autoencoders are usually trained one layer at a time, even though transformer residual stream activations are strongly coupled across depth. This creates a practical problem…
Residualized Temporal Sparse Autoencoders for Interpreting Diffusion Models
Calvin Yeung, Prathyush Poduval, Ali Zakeri +2
Text-to-image diffusion models generate images through an iterative denoising process, so internal neural layers produce trajectories of activations rather than single static repre…
Vestigial pairing from fluctuating magnetism and triplet superconductivity
Yanek Verghis, Denis Sedov, Jakob WeÃling +2
We study the finite-temperature vestigial superconducting phases of a two-dimensional system of fluctuating spin-triplet pairing and spin magnetism. Denoting the respective primary…
Enabling Group Fairness in Graph Unlearning via Bi-level Debiasing
Yezi Liu, Prathyush Poduval, Wenjun Huang +3
Graph unlearning is a crucial approach for protecting user privacy by erasing the influence of user data on trained graph models. Recent developments in graph unlearning methods ha…
Explainable Differential Privacy-Hyperdimensional Computing for Balancing Privacy and Transparency in Additive Manufacturing Monitoring
Fardin Jalil Piran, Prathyush P. Poduval, Hamza Errahmouni Barkam +2
Machine Learning (ML) models integrated with in-situ sensing offer transformative solutions for defect detection in Additive Manufacturing (AM), but this integration brings critica…