[News item updated on 2 May]
- "Layer-wise Quantization for Quantized Optimistic Dual Averaging"
Authors: Anh Duc Nguyen, Ilia Markov, Zhengqing Wu, Ali Ramezani-Kebrya, Kimon Antonakopoulos, Dan Alistarh, and Volkan Cevher
We develop a general layer-wise quantization framework with tight variance and code-length bounds, adapting to the heterogeneities over the course of training. We empirically achieve up to a 150% speedup over the baselines in end-to-end training time for training Wasserstein GAN on 12+ GPUs.
- "Self-Organizing Visual Prototypes for Non-Parametric Representation Learning"
Joint work with Thalles Silva, Helio Pedrini, Adín Ramírez Rivera
We propose a new prototypical self-supervised method that leverages context from the representations to enrich them when performing comparisons. We comprehensively assess the new learned representations on benchmarks such as retrieval, linear evaluation, fine-tuning, and object detection. The proposal achieves the best performance on many retrieval bench marks and shows increased performance gains as more complex encoders are utilized.
-
"DocVXQA: Context-Aware Visual Explanations for Document Question Answering"
Mohamed Ali Souibgui, Changkyu Choi*, Andrey Barsky, Kangsoo Jung, Ernest Valveny, Dimosthenis Karatzas
We propose DocVXQA, a framework for visually self-explainable document question answering that produces accurate answers while generating visual heatmaps for interpretability. By encoding explainability as learning criteria, DocVXQA balances performance and trust through context-aware explanations.
* Changkyu Choi is a Visual Intelligence postdoctoral research fellow from University of Tromsø, who has a long-term residence in our research group.
See you at ICML 2025!