MBZUAI ·
A Mixture of Experts (MoE) layer is a sparsely activated deep learning layer. It uses a router network to direct each token to one of the experts. Yuanzhi Li, an assistant professor at CMU and affiliated faculty at MBZUAI, researches deep learning theory and NLP. Why it matters: This highlights MBZUAI's engagement with cutting-edge deep learning research, specifically in efficient model design.
arXiv ·
MBZUAI researchers introduce BiMediX, a bilingual (English and Arabic) mixture of experts LLM for medical applications. The model is trained on BiMed1.3M, a new 1.3 million bilingual instruction dataset and outperforms existing models like Med42 and Jais-30B on medical benchmarks. Code and models are available on Github.
arXiv ·
This paper introduces Adaptive Entropy-aware Optimization (AEO), a new framework to tackle Multimodal Open-set Test-time Adaptation (MM-OSTTA). AEO uses Unknown-aware Adaptive Entropy Optimization (UAE) and Adaptive Modality Prediction Discrepancy Optimization (AMP) to distinguish unknown class samples during online adaptation by amplifying the entropy difference between known and unknown samples. The study establishes a new benchmark derived from existing datasets with five modalities and evaluates AEO's performance across various domain shift scenarios, demonstrating its effectiveness in long-term and continual MM-OSTTA settings.
MBZUAI ·
A recent study questions the necessity of deep ensembles, which improve accuracy and match larger models. The study demonstrates that ensemble diversity does not meaningfully improve uncertainty quantification on out-of-distribution data. It also reveals that the out-of-distribution performance of ensembles is strongly determined by their in-distribution performance. Why it matters: The findings suggest that larger, single neural networks can replicate the benefits of deep ensembles, potentially simplifying model deployment and reducing computational costs in the region.
arXiv ·
The study compares deep learning models trained via transfer learning from ImageNet (TII-models) against those trained solely on medical images (LMI-models) for disease segmentation. Results show that combining outputs from both model types can improve segmentation performance by up to 10% in certain scenarios. A repository of models, code, and over 10,000 medical images is available on GitHub to facilitate further research.
arXiv ·
Researchers at MBZUAI introduce FissionFusion, a hierarchical model merging approach to improve medical image analysis performance. The method uses local and global aggregation of models based on hyperparameter configurations, along with a cyclical learning rate scheduler for efficient model generation. Experiments show FissionFusion outperforms standard model souping by approximately 6% on HAM10000 and CheXpert datasets and improves OOD performance.
arXiv ·
The authors introduce Nile-Chat, a collection of LLMs (4B, 3x4B-A6B, and 12B) specifically for the Egyptian dialect, capable of understanding and generating text in both Arabic and Latin scripts. A novel language adaptation approach using the Branch-Train-MiX strategy is used to merge script-specialized experts into a single MoE model. Nile-Chat models outperform multilingual and Arabic LLMs like LLaMa, Jais, and ALLaM on newly introduced Egyptian benchmarks, with the 12B model achieving a 14.4% performance gain over Qwen2.5-14B-Instruct on Latin-script benchmarks; all resources are publicly available. Why it matters: This work addresses the overlooked aspect of adapting LLMs to dual-script languages, providing a methodology for creating more inclusive and representative language models in the Arabic-speaking world.