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.
MBZUAI Professor Fakhri Karray and co-authors from the University of Waterloo have published "Elements of Dimensionality Reduction and Manifold Learning," a textbook on methods for extracting useful components from large datasets. The book addresses the challenge of the "curse of dimensionality," where growth in datasets complicates their use in machine learning. Karray developed the material from a popular course he taught at Waterloo. Why it matters: The textbook provides a unified resource for students and researchers in machine learning and AI, addressing a foundational challenge in processing high-dimensional data, relevant to diverse applications in the region.
A talk introduces a computational framework for learning a compact structured representation for real-world datasets, that is both discriminative and generative. It proposes to learn a closed-loop transcription between the distribution of a high-dimensional multi-class dataset and an arrangement of multiple independent subspaces, known as a linear discriminative representation (LDR). The optimality of the closed-loop transcription can be characterized in closed-form by an information-theoretic measure known as the rate reduction. Why it matters: The framework unifies concepts and benefits of auto-encoding and GAN and generalizes them to the settings of learning a both discriminative and generative representation for multi-class visual data.
KAUST Professor Marc Genton and his former postdoc Stefano Castruccio jointly won the 2017 Wilcoxon Award for their paper in Technometrics. Their paper, "Compressing an ensemble with statistical models: An algorithm for global 3D spatio-temporal temperature," details a data-compression scheme for climate simulations. The method reduces data-storage requirements and accelerates climate research capacity. Why it matters: This award highlights KAUST's contribution to statistical methods for climate modeling and big data analysis, particularly relevant for studying renewable energy resources in Saudi Arabia.
A CMU researcher, Dr. Hongyi Wang, presented an evaluation of gradient compression methods in distributed training, finding limited speedup in most realistic setups. The research identifies the root causes and proposes desirable properties for gradient compression methods to provide significant speedup. The talk was promoted by MBZUAI. Why it matters: Understanding the limitations of gradient compression techniques can help optimize distributed training strategies for AI models in the region.
Technology Innovation Institute's (TII) Directed Energy Research Center (DERC) is integrating machine learning (ML) techniques into signal processing to accelerate research. One project used convolutional neural networks to predict COVID-19 pneumonia from chest x-rays with 97.5% accuracy. DERC researchers also demonstrated that ML-based signal and image processing can retrieve up to 68% of text information from electromagnetic emanations. Why it matters: This adoption of ML for signal processing at TII highlights the potential for advanced AI techniques to enhance research and security applications in the UAE.
Researchers at MBZUAI release SlimPajama-DC, an empirical analysis of data combinations for pretraining LLMs using the SlimPajama dataset. The study examines the impact of global vs. local deduplication and the proportions of highly-deduplicated multi-source datasets. Results show that increased data diversity after global deduplication is crucial, with the best configuration outperforming models trained on RedPajama.
This paper benchmarks multilingual and monolingual LLM performance across Arabic, English, and Indic languages, examining model compression effects like pruning and quantization. Multilingual models outperform language-specific counterparts, demonstrating cross-lingual transfer. Quantization maintains accuracy while promoting efficiency, but aggressive pruning compromises performance, particularly in larger models. Why it matters: The findings highlight strategies for scalable and fair multilingual NLP, addressing hallucination and generalization errors in low-resource languages.