- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes GradPath, a parameter-efficient fine-tuning method for medical image analysis. Each hidden unit of a feed-forward block is treated as a "pathway" made of one row of the input projection and one column of the output projection. A score, the absolute product of the unit's activation and the loss gradient averaged over the training set, ranks the units, and the paired row and column of the top K units in every block are updated in place, without new modules. A variant, GradPath-S, adds gated depthwise convolutions with three kernel sizes on the selected channels. With ViT-B on seven classification datasets, GradPath and GradPath-S reach 90.96% and 91.55% average accuracy, against 90.77% for GPS and 90.70% for Adapter. Further experiments cover Swin-T and ConvNeXt-T, five segmentation datasets with DINOv3, few-shot adaptation of BiomedCLIP on seven datasets, reduced training-data ratios and VTAB-1K. The paper also reports lower training memory, FLOPs and time per iteration than full fine-tuning, GPS and DKA, and argues that the selected pathways are task-specific.
Critical strengths
- The method is simple and clean. A hidden unit with its input row and output column is a natural adaptation unit, the pretrained weights are updated in place, and no module is added, so the inference graph is unchanged and the idea carries over to Swin and ConvNeXt.
- The evaluation is broad: classification on seven datasets and three backbones, segmentation, few-shot adaptation of a vision-language model, reduced-data settings and VTAB-1K, with analyses of block-wise importance and of transferring selected pathways between tasks.
- Efficiency is measured as peak memory, FLOPs and time per iteration with a forward/backward breakdown, and not only as a parameter count.
Critical weaknesses
- The evidence for the two central ideas, pairing and task-sensitive selection, is thin. In Table 5, the four alternative selection or update schemes differ from GradPath by at most 0.7 points on three datasets, and on BTMRI the independent variant gives the same accuracy. Randomly selected pathways are not tested in the fully supervised setting. In Fig. 4, a Jaccard overlap of 0.18 to 0.27 is called low, but two random sets of 64 out of 3072 units would overlap by about 0.01, so a third or more of the selected pathways are shared between unrelated tasks. On three of the four datasets, pathways selected on other tasks recover 60 to 76% of the gain of target-specific over random pathways. The score itself is the first-order Taylor importance known from pruning (Molchanov et al., 2017), and structured sparse fine-tuning that selects FFN channels and updates pretrained weights in place (S2FT, Yang et al., 2024) is not discussed. The theory adds little: Proposition 1 follows from the definition of the score, and Appendix D holds for any fixed subset of parameters.
- The paper does not mention that the released training script first trains the classifier head in a warm-up stage before the pathway scores are computed. It is not stated whether the baselines were trained in the same way. The released scripts for GradPath and GradPath-S also differ in more than the spatial module (64 versus 32 pathways, no bias update versus bias update, one versus five warm-up epochs), so the gain attributed to the spatial enhancement may be confounded. Trainable parameters are not reported for Tables 1 and 4, and the tokens used for scoring are not specified. The efficiency comparison in Fig. 5 covers only full fine-tuning, DKA and GPS, leaves out Adapter, LoRA and SSF, and is measured for GradPath, while the best accuracies come from GradPath-S.
Questions for authors
- Were the baselines trained with the same protocol as in the released code (classifier warm-up, weight decay, validation-driven schedule, early stopping), and were learning rates tuned per method? Why are LoRA and VPT below linear probing?
- Where do the SSF and GPS rows in Table 9 come from? The GPS paper reports, for example, 81.1 on CIFAR-100 and 84.0 on dSprites-Loc, against 71.8 and 71.8 here.
- Is the classifier trained before the pathway scores are computed, and how is this handled for the randomly initialised segmentation decoder? Which settings (K, bias update, warm-up) were used for GradPath-S in Table 1, and what does GradPath-S give with the settings of GradPath?
- What is the accuracy with randomly selected paired pathways in the setting of Table 5? Why do the target-specific accuracies in Fig. 4(a) (79.3, 82.7, 73.1, 79.6) differ from the 8-shot GradPath results in Table 4 (82.0, 81.6, 75.2, 80.1)?
- The dataset splits follow the 11-dataset benchmark of BiomedCoOp. Why are different subsets of seven used for classification and for few-shot learning, and why is BUSI not used?
2 comments
Thank you
Thank you for your thoughtful reassessment and encouraging feedback. We are glad that our responses have addressed your concerns. Your constructive comments have helped us improve the clarity and completeness of the paper.
Best regards, The Authors
Thank you for the thorough and extensive rebuttal. You have addressed all of my concerns to my satisfaction. Accordingly, I will raise my recommendation from 2 to 3, as I now have considerably greater confidence in the paper's contribution and merits.