Improving Knowledge Distillation in Transfer Learning with Layer-wise Learning Rates

TL;DR

This paper introduces a layer-wise learning rate adjustment scheme based on Jacobian, attention, and Hessian differences to enhance transfer learning, especially for complex tasks, improving performance and stability.

Contribution

The work proposes a novel layer-wise learning scheme that adjusts parameters based on Jacobian, attention, and Hessian differences, improving transfer learning performance.

Findings

Enhanced learning stability across datasets

Performance gains increase with task complexity

Effective for attention and derivative-based transfer methods

Abstract

Transfer learning methods start performing poorly when the complexity of the learning task is increased. Most of these methods calculate the cumulative differences of all the matched features and then use them to back-propagate that loss through all the layers. Contrary to these methods, in this work, we propose a novel layer-wise learning scheme that adjusts learning parameters per layer as a function of the differences in the Jacobian/Attention/Hessian of the output activations w.r.t. the network parameters. We applied this novel scheme for attention map-based and derivative-based (first and second order) transfer learning methods. We received improved learning performance and stability against a wide range of datasets. From extensive experimental evaluation, we observed that the performance boost achieved by our method becomes more significant with the increasing difficulty of the learning task.