Towards Lightweight Adaptation of Massive Neural Network Models

The advent of foundation models has revolutionized artificial intelligence, enabling unprecedented performance across natural language processing, computer vision, and multimodal tasks. However, the sheer scale of these models, often encompassing billions or even trillions of parameters, poses significant challenges for adaptation to downstream tasks, including high computational cost, memory constraints, and environmental impact. Parameter-Efficient Fine-Tuning (PEFT) has emerged as a principled solution to this problem, providing mechanisms to adapt large pretrained models by updating only a small subset of parameters while keeping the majority of the model frozen. This survey provides a comprehensive examination of PEFT techniques, categorizing them into adapters, low-rank adaptation methods, prompt-and prefix-based approaches, and hybrid strategies, and highlighting the theoretical foundations that justify their efficiency and effectiveness. We discuss the structural design choices underlying these methods, their empirical performance across a range of tasks and domains, and the trade-offs between parameter efficiency, inference cost, generalization, and robustness. The survey further explores applications of PEFT in natural language understanding, computer vision, multimodal learning, continual adaptation, and personalization, illustrating how modular and low-dimensional updates enable scalable and sustainable AI. Additionally, we address challenges related to evaluation metrics, robustness under distribution shift, fairness, interpretability, and deployment constraints, emphasizing the need for standardized benchmarks and multi-objective assessments. Interdisciplinary opportunities, including the integration of PEFT with reinforcement learning, federated learning, and neuroscience-inspired mechanisms, are examined as promising directions for future research. Finally, we discuss the broader societal and environmental implications of PEFT, highlighting its potential to democratize access to large models, reduce energy consumption, and promote ethical, equitable, and sustainable AI practices. By synthesizing current knowledge, identifying open challenges, and proposing avenues for innovation, this survey aims to provide researchers and practitioners with a definitive reference for understanding, implementing, and advancing parameter-efficient adaptation of foundation models. The findings suggest that PEFT not only addresses pressing efficiency constraints but also serves as a central paradigm for responsible, scalable, and adaptable artificial intelligence in the era of ever-growing foundation models.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.