Improving Reliability of Fine-tuning with Block-wise Optimisation

Finetuning can be used to tackle domain specific tasks by transferring knowledge learned from pre-trained models. However, previous studies on finetuning focused on adapting only the weights of a task-specific classifier or re-optimising all layers of the pre-trained model using the new task data. The first type of method cannot mitigate the mismatch between a pre-trained model and the new task data, and the second type of method easily causes over-fitting when processing tasks with limited data. To explore the effectiveness of fine-tuning, we propose a novel block-wise optimisation mechanism, which adapts the weights of a group of layers of a pre-trained model. This work presents a theoretical framework and empirical evaluation of block-wise fine-tuning to find a reliable fine tuning strategy. The proposed approach is evaluated on two datasets, Oxford Flowers (OXF) and Caltech 101 (CAL), using 15 commonly used pre-trained base models. Results indicate that the proposed strategy consistently outperforms the baselines in terms of classification accuracy, although the specific block leading to optimal performance may vary across models. The investigation reveals that selecting a block from the fourth quarter of a base model generally yields improved performance compared to the baselines. Overall, the block-wise approach consistently outperforms the baselines and exhibits higher accuracy and reliability. This study provides valuable insights into the selection of salient blocks and highlights the effectiveness of block-wise fine-tuning in achieving improved classification accuracy in various models and datasets