标题： 临界动力学支配深度学习 显示英文标题

标题： Critical dynamics governs deep learning

摘要： 人工智能通过更大和更深的神经网络迅速发展，但如何优化网络动态以实现性能和适应性的基本问题仍然存在。 本研究显示，深度神经网络（DNNs）在接近临界相变的状态下表现最佳，这种状态介于活跃和不活跃的动力学之间，类似于生物大脑。 结合物理学和神经科学，我们证明临界性提供了一个统一的原则，将DNNs的结构、动态和功能联系起来。 分析了超过80个最先进的模型，我们首先报告称，过去十年中准确性的提高与隐式地向更临界动力学演进同时发生。 架构和训练创新无意中引导网络向这一最佳状态发展。 其次，基于这些见解，我们开发了一种显式驱动网络达到临界性的训练方法，提高了鲁棒性和性能。 第三，我们表明人工智能中的基本问题，包括深度持续学习中性能的丧失，是由临界性的丧失引起的，而保持临界性可以恢复性能。 这项工作通过强调动态优化与规模并重，将临界性引入人工智能发展的基础框架中。 它将人工智能与物理学和生物皮层网络功能联系起来，激发了DNNs中的新颖自调制策略。 这些发现为设计高效、可适应和高性能的人工智能系统提供了理论上有根据的前进路径，灵感来源于生物神经系统的原理。 显示英文摘要

摘要： Artificial intelligence has advanced rapidly through larger and deeper neural networks, yet fundamental questions remain about how to optimize network dynamics for performance and adaptability. This study shows that deep neural networks (DNNs), like biological brains, perform optimally when operating near a critical phase transition - poised between active and inactive dynamics. Drawing from physics and neuroscience, we demonstrate that criticality provides a unifying principle linking structure, dynamics, and function in DNNs. Analyzing more than 80 state-of-the-art models, we first report that improvements in accuracy over the past decade coincided with an implicit evolution toward more critical dynamics. Architectural and training innovations unknowingly guided networks toward this optimal regime. Second, building on these insights, we develop a training method that explicitly drives networks to criticality, improving robustness and performance. Third, we show that fundamental problems in AI, including loss of performance in deep continual learning, are caused by loss of criticality and that maintaining criticality rescues performance. This work introduces criticality as a fundamental framework for AI development by emphasizing dynamic optimization alongside scale. It bridges artificial intelligence with physics and biological cortical network function inspiring novel self-tuning strategies in DNNs. The findings offer a theoretically grounded path forward in designing efficient, adaptable, and high-performing artificial intelligence systems drawing inspiration from principles observed in biological neural systems.