Harnessing Orthogonality to Train Low-Rank Neural Networks

Abstract

This study explores the learning dynamics of neural networks by analyzing the singular value decomposition (SVD) of their weights throughout training. Our investigation reveals that an orthogonal basis within each multidimensional weight's SVD representation stabilizes during training. Building upon this, we introduce Orthogonality-Informed Adaptive Low-Rank (OIALR) training, a novel training method exploiting the intrinsic orthogonality of neural networks. OIALR seamlessly integrates into existing training workflows with minimal accuracy loss, as demonstrated by benchmarking on various datasets and well-established network architectures. With appropriate hyperparameter tuning, OIALR can surpass conventional training setups, including those of state-of-the-art models.

References in corpus (14)

• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• Escaping the Big Data Paradigm with Compact Transformers
• Revisiting ResNets: Improved Training and Scaling Strategies
• Rigging the Lottery: Making All Tickets Winners
• Structured Pruning of Large Language Models
• Language model compression with weighted low-rank factorization
• Directional convergence and alignment in deep learning
• Massively Parallel Genetic Optimization through Asynchronous Propagation of Populations
• Low-rank lottery tickets: finding efficient low-rank neural networks via matrix differential equations
• kDecay: Just adding k-decay items on Learning-Rate Schedule to improve Neural Networks
• Principal Component Networks: Parameter Reduction Early in Training
• Greenformer: Factorization Toolkit for Efficient Deep Neural Networks
• Bort: Towards Explainable Neural Networks with Bounded Orthogonal Constraint
• Adaptive Low-Rank Factorization to regularize shallow and deep neural networks