Research on efficient transformer models — knowledge distillation, multimodal vision–language learning, and the quality of training data.
Accuracy alone hides how models forget in continual learning. We introduce six softmax-derived metrics covering rank, confidence, and distributional divergence that expose class-level forgetting patterns invisible to accuracy. Using these as loss weights or replay sampling criteria reduces forgetting by up to 7.7 pp on TinyImageNet over uniform experience replay.
CutMix assigns labels by patch area, not by visible object content, a systematic bias that mislabels 21.5% of samples and creates ghost labels in 17%. OA-CutMix replaces the label with one derived from object area, leaving the image mixing unchanged. It matches or beats 10+ static and dynamic mixing methods across 4 architectures and 6 datasets.
We show that newer text-to-image models are progressively worse as training data generators, despite better visual quality, because they collapse to a narrow aesthetic-centric distribution that diverges from real data.
We use a frozen text encoder on image captions as a lightweight training-time auxiliary objective for image classifiers. The text components are dropped at inference, leaving a fast, unimodal vision model. Accuracy on ImageNet improves by up to +2.7 p.p. and downstream transfer by +1.0 p.p. on average, outperforming vision knowledge distillation at a fraction of the compute.
We fix Forward-Forward's slow inference by replacing per-class passes with a single forward pass through hyperspherical prototype matching. Thus, we achieve 40× faster inference with competitive accuracy.
We introduce PRISM, a framework that disentangles architectural priors for dataset distillation, outperforming single-teacher setups.
We introduce SubZeroCore, a novel, training-free coreset selection method that integrates submodular coverage and density into a single, unified objective.
We present HyperCore, a lightweight adaptive coreset selection framework designed for noisy environments. HyperCore utilizes per class hypersphere models and adaptively selects pruning thresholds.
We extend pretrained super-resolution models to larger images by using local-aware prompts.
We improve dataset distillation by distilling only a representative coreset.
This paper proposes a new method to parameterize open loop controls in stochastic optimal control problems using path signatures. We show that these controls are dense in the space of all admissible controls and establish conditions for stability of the controlled dynamics and target functional.
We improve the training of vision transformers by segmenting and recombining objects and backgrounds from datasets. This makes the transformers more accurate, as well as more robust.
A comprehensive benchmark and analysis of more than 45 transformer models for image classification to evaluate their efficiency, considering various performance metrics. We find the optimal architectures to use and uncover that model-scaling is more efficient than image scaling.
We conduct the first systematic study of dataset distillation for Super-Resolution.
This paper introduces TaylorShift, a novel reformulation of the attention mechanism using Taylor softmax that enables computing full token-to-token interactions in linear time. We analytically and empirically determine the crossover points where employing TaylorShift becomes more efficient than traditional attention. TaylorShift outperforms the traditional transformer architecture in 4 out of 5 tasks.
We speed up diffusion classifiers by utilizing a label hierarchy and pruning unrelated paths.
We utilize the TaylorShift attention mechanism for global pixel-wise-attention in image super-resolution.
We consider a stochastic control problem and try to solve it using the signature method.
We extend and test KEdge, an interpretable-by-design approach for graph neural networks, and compare it to gradient-based attribution techniques.