Bayesian fine-tuning method improves model accuracy and confidence estimates
Geometry-Aware Bayesian Parameter-Efficient Fine-Tuning on the Stiefel Manifold via Stein Variational Gradient Descent
Machine Learning
Summary
Tuning large pretrained models efficiently is tricky because most parameters are fixed. The authors present a way to fine-tune part of these models using a math technique that respects geometric constraints. This technique produces multiple possible tuned models that better estimate their own confidence in predictions. As a result, the tuned models are both more accurate and better at knowing when they might be wrong.
What this means in practice
- •For machine learning engineers: Produce pretrained model adapters that better estimate prediction uncertainty while improving accuracy in resource-limited fine-tuning scenarios.
- •For autonomous vehicle developers: Generate better-calibrated models that can express uncertainty, enhancing safety-critical decision making in changing environments.
Authors
Quang-Duy Tran, Trung Le, Bao Duong, Phuoc Nguyen, Thin Nguyen
Abstract
Several geometry-aware approaches to low-rank adaptation have emerged for parameter-efficient fine-tuning of large pre-trained models. These methods aim to take full advantage of the geometric structure of low-rank manifolds for improving the efficiency in subspace utilization and reducing redundancy by enforcing orthogonality constraints during optimization. The strong empirical results of these techniques have motivated further study into whether predictions from such geometry-based adaptation methods could be overconfident. In this paper, we build on the singular value decomposition factorization of adapters to develop a framework based on Stein variational gradient descent (SVGD). In this formulation, the low-rank matrices are transported along the Stiefel manifold to match the targeted distributions while retaining their crucial geometric structure. Since this geometry-aware SVGD approach provides multiple solutions during inference, it supports uncertainty quantification and produces better-calibrated adapters on the Stiefel manifold. Extensive experiments show that our method delivers strong model calibration and attains higher prediction accuracy than SVGD and related uncertainty estimation methods that are formulated in Euclidean space.