arXiv:2609.35288 · Preprint
Anti-collapse objectives applied after the projector can leave the backbone low-rank. SACReg, motivated by an analysis of λ-balance in linear networks, regularizes the backbone directly; λ-JEPA improves over LeJEPA and VISReg on ImageNet-1k and over V-JEPA 2 on video benchmarks.
Advances in Neural Information Processing Systems · NeurIPS 2026
A post-hoc, per-sample trust score for conditional generations that needs only the training distribution: global realism plus attribute-wise faithfulness. It supports filtering, ranking, and abstention, even before full decoding.
Advances in Neural Information Processing Systems · NeurIPS 2026
A diffusion model for fluorescent microscopy that generates all Cell Painting channels jointly, across cell types and perturbations, with an alignment loss to OpenPhenom embeddings that preserves organelle-level structure.
Advances in Neural Information Processing Systems · NeurIPS 2025
ORA scores a feature by the angle between it and its projection onto the decision boundaries, measured from the mean of in-distribution features. It is model-agnostic, hyperparameter-free, scale-invariant, and lets scores from several models be summed.
arXiv:2410.04499 · Preprint
A modular approach to performative label shift for pretrained backbones: it adapts a model before the shift and anticipates how different models would shift the distribution, enabling informed model selection and zero-shot transfer across backbone updates.
IEEE International Conference on Image Processing · ICIP 2024
Group activity recognition from a single RGB frame, with an attention-pooling visual branch and a coordinate branch over actor boxes, plus a systematic re-annotation of the Volleyball dataset.
arXiv:2308.06624 · Preprint
Domain generalization that keeps domain-specific features alongside domain-invariant ones through an additive disentanglement, with a latent-space remix augmentation that mixes samples across domains.
IEEE Signal Processing and Communications Applications Conference · SIU 2020
Node rankings from degree, closeness centrality, clustering coefficient, and PageRank guide greedy coloring to substantially better colorings; an RL extension with a parallel C++ backend.
A re-implementation of Andrej Karpathy's nanoGPT, designed to be simpler and easier to update, evaluated on Tiny Shakespeare.
Uniform Manifold Approximation and Projection implemented from scratch, with experiments on MNIST and Load Digits.
Course project
Parallel implementations of direction-optimized BFS on CPU (OpenMP) and GPU (CUDA).
No publications match. Try a different search or filter.