A Hessian-informed hyperparameter optimization for differential learning rate
This work addresses hyperparameter optimization for differential learning rates, which is incremental as it builds on existing techniques like parameter-efficient fine-tuning.
The paper tackles the problem of optimizing differential learning rates by proposing Hessian-informed differential learning rate (Hi-DLR), which adaptively captures loss curvature to dynamically set learning rates, empirically improving convergence.
Differential learning rate (DLR), a technique that applies different learning rates to different model parameters, has been widely used in deep learning and achieved empirical success via its various forms. For example, parameter-efficient fine-tuning (PEFT) applies zero learning rates to most parameters so as to significantly save the computational cost. At the core, DLR leverages the observation that different parameters can have different loss curvature, which is hard to characterize in general. We propose the Hessian-informed differential learning rate (Hi-DLR), an efficient approach that solves the hyperparameter optimization (HPO) of learning rates and captures the loss curvature for any model and optimizer adaptively. Given a proper grouping of parameters, we empirically demonstrate that Hi-DLR can improve the convergence by dynamically determining the learning rates during the training.