Posts

We present a unified framework that combines expressive neural embeddings with interpretable multi-output Gaussian processes for single-cell genomics. Joint representations of cells and genes reveal meaningful links between cell clusters and their marker genes via an interpretable gene-relevance map.

We introduce ModelAuditor, a self-reflective agent that simulates clinically relevant distribution shifts and produces interpretable reports on likely failure modes. Across multiple medical imaging domains, it recovers up to 25% of performance lost under shift while providing actionable deployment insights.

We present the first systematic framework for application-driven validation of posterior-based methods in inverse problems. Adapting concepts from object detection enables mode-centric validation with interpretable, application-focused metrics, demonstrated on multiple medical imaging use cases.

We introduce LVLM-Aided Visual Alignment (LVLM-VA), which aligns small vision models with human domain knowledge using large vision-language models. A bidirectional interface translates model behavior into natural language and expert instructions into image-level critiques, improving performance while reducing fine-grained feedback needs.

We analyze how predictive uncertainty can guide memory management to mitigate catastrophic forgetting and introduce a generalized-variance–based uncertainty measure. Uncertainty-aware sampling improves retention across tasks. Published in the Journal of Transactions on Machine Learning Research.

We propose IUPM, a label-free method for tracking performance under gradual distribution shifts using optimal transport. IUPM quantifies uncertainty in its estimates and guides targeted labeling to restore reliability, outperforming existing baselines across scenarios.

We propose an uncertainty-aware memory-based approach for online Federated Continual Learning. Using a Bregman Information estimator to guide selective replay, the method reduces catastrophic forgetting across modalities while preserving privacy and communication efficiency.

We introduce an unsupervised framework to detect and mitigate shortcut learning in transformers. The method improves both worst-group and average accuracy while reducing annotation effort, offering interpretable insights for experts and running efficiently on consumer hardware.

We introduce GEMNA, a deep learning framework for mass spectrometry–based metabolomics that uses graph and edge embeddings with anomaly detection. GEMNA outperforms traditional tools in untargeted studies, producing clearer clusters and improved biological insights.

Our latest work presents a new approach to disentangle image generation performance by decomposing cosine similarity into cluster-level contributions using central kernel alignment. This allows us to quantify how different pixel regions contribute to overall image quality, enabling more fine-grained evaluation and improved explainability of generative models across real-world use cases.