arXiv ·
Researchers at MBZUAI have introduced MedMerge, a transfer learning technique that merges weights from independently initialized models to improve performance on medical imaging tasks. MedMerge learns kernel-level weights to combine features from different models into a single model. Experiments across various medical imaging tasks demonstrated performance gains of up to 7% in F1 score.
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 paper introduces Sparse-Quantized Representation (SpQR), a new compression format and quantization technique for large language models (LLMs). SpQR identifies outlier weights and stores them in higher precision while compressing the remaining weights to 3-4 bits. The method achieves less than 1% accuracy loss in perplexity for LLaMA and Falcon LLMs and enables a 33B parameter LLM to run on a single 24GB consumer GPU. Why it matters: This enables near-lossless compression of LLMs, making powerful models accessible on resource-constrained devices and accelerating inference without significant accuracy degradation.