- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes GIRA, a source-free domain adaptation framework for sequence regression. It uses generalized classes and a conditional generative model to construct adaptation signals from unlabeled target data. Experiments on three tasks and nine transfer scenarios show improvements over the zero-shot models and several adaptation baselines.
Critical strengths
-
The paper addresses an problem in source-free domain adaptation for sequence regression. This is a meaningful research direction with potential practical value.
-
The evaluation covers three different tasks and nine transfer scenarios, providing evidence that the approach can work across different regression settings. The ablation studies also offer some insight into the roles of the proposed components.
Critical weaknesses
-
The main adaptation loss is applied to generated inputs rather than directly to real target inputs. The current model prediction also serves as the generator's conditioning input, and gradients propagate through this path. It is therefore unclear whether minimizing the loss improves predictions on real target data or mainly adjusts the conditioning and generated samples.
-
The theoretical analysis relies on several assumptions, and the resulting risk bound depends on terms such as the generalized-class error and generated–target discrepancy, which are difficult to assess in the source-free setting. Although the appendix derives further bounds, it remains unclear how informative these bounds are under the assumptions required and how tightly they relate to the actual target risk.
Questions for authors
-
How does conditioning the generator on the current prediction , rather than the fixed initial prediction , affect adaptation? Is there evidence that the gradient through the conditioning path is important?
-
The GC-only ablation performs substantially worse than the zero-shot model on T#4. Which components are primarily responsible for preventing negative transfer, and how much does the generative inverse mapping contribute beyond regularization and filtering?
-
GIRA performs less favorably in speech enhancement under larger domain shifts. How sensitive is the method to violations of the assumption that the pretrained model retains basic predictive capability on the target domain?
3 comments
Responses to Author
Thank you for the detailed response and additional experiments.
Responses to Reviewer Cxpt
Thank you for taking the time to consider our response and additional experiments. Your questions helped us explain the role of GIRA’s two gradient paths more clearly.
Responses to Reviewer Cxpt
We thank the reviewer for the constructive review and for recognizing our research direction, evaluation and ablation studies. We respond to the concerns below.
C1. Does a loss on generated inputs improve real target predictions, or only adjust the conditioning?
We thank the reviewer for raising this point. Minimizing Eq. (12) improves the predictions on real target data, not only the conditioning or the generated samples. G is frozen in Step 2, so N is the only trained network, and every change of the condition or the generated samples results from a change of N.
This works like self-supervised pseudo-labeling in classification SFDA (e.g., SHOT), where the correctly labeled majority of a class corrects its mislabeled minority. G fills each generalized class, identified by the source prediction of one real input, with many target-style inputs paired with that identifier. Most pairs are correctly assigned, so fitting N to them corrects its predictions on the target inputs they resemble, including the real ones (see gm9L-C1).
Unlike classification, where the classes are given, the GCs are defined by the model itself, so the position of an identifier can also be wrong. Conditioning G on the current prediction N(xT) lets the class position move with N while the target yˉ(xT) stays fixed. Its gradient takes two paths. The shallow path, through the outer N, corrects the wrongly assigned generated inputs. The deep path, through G back to the inner N(xT), moves the class position to where the generated population is consistent with its identifier. Since the inner N(xT) is the prediction on the real input xT, this path adjusts the real-target predictions directly.
C2. How informative is the risk bound?
The bound is informative because its terms track the actual target risk. In its expanded form (Eq. 48), all data-dependent terms except the error–uncertainty envelope of Eq. (25) are label-free, namely LA, the generator Score, and the UTF threshold and coverage. The generator Score ranks the ten T#4 generator checkpoints exactly as the adapted MAE (Table 5). As the bound predicts, GIRA improves target accuracy once the generator is trained for 3,000 of 10,000 steps and the source model for 15 of 50 epochs (Tables 5, 17), i.e., when most generated pairs are correctly assigned (see Ru2P-C1). Like other domain adaptation bounds, it identifies which quantities control the risk rather than giving a numerical estimate.
C3. Does conditioning on N(xT) rather than the fixed yˉ(xT), with its gradient path, matter?
We thank the reviewer for this question and ran this ablation on T#1, T#4 and T#5 (gm9L-C3, App. C.6). Conditioning on the fixed yˉ(xT) or stopping the gradient through N(xT) still improves over zero-shot, while the full gradient path gives the best MAE. Both variants keep only the shallow path, which corrects the wrongly assigned generated inputs and provides most of the gain. The deep path adds the rest by correcting the class positions, i.e., the predictions N(xT) on the real inputs (see C1), so the gradient through the condition is important.
C4. Why does GC-only fail on T#4, and what does the inverse contribute beyond regularization and filtering?
We appreciate this question. In Table 4, LSCC is what prevents negative transfer on T#4. Without it, MAE is 1592.41, close to GC-only (1603.67), whereas without LREG (362.85) or UTF (313.12) the model still beats zero-shot (498.13). GC-only fails through a feature-scale drift that shifts the whole output by a constant offset (Fig. 3b). It is strongest on T#4, whose domains differ most (Fig. 4a).
The gain itself comes from the inverse mapping. With UTF, LREG and LSCC kept identical, removing LGC or replacing it by Gaussian-perturbation consistency or UTF self-training leaves MAE within 3% of zero-shot (see Ru2P-C2).
C5. How sensitive is GIRA to violating the basic-performance assumption?
App. C.4 (Table 17) varies source quality on T#4 with source checkpoints from epochs 10 to 50. At the weakest checkpoint, whose zero-shot MAE (910.24) is 83% higher than that of the final one, MAE worsens by only 1.63%. Gains are positive from epoch 15 (+11.68%) and grow to +47.96% at epoch 50.
In speech enhancement, T#9 (J+O→V+M) has a large JSUT–VCTK shift (Fig. 4b), yet GIRA still improves on zero-shot (MAE 10.84→10.61, SegSNR 13.22→13.46 dB). It also has the lowest MAE among source-free methods, and only the source-using CDAN and DANN reach a lower MAE.
In short, GIRA remains beneficial under moderate violations of the assumption and degrades only slightly under a strong one.
Related Revisions