- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
FedVFG addresses federated classification with both domain shift and heterogeneous client architectures. Clients upload class means and diagonal variances. The server normalizes these statistics by a client RMS feature scale and aggregates them by domain and class. Clients interpolate statistics from two supported domains, transform local features through a bounded location-scale map, and train with cross-entropy on both original and virtual features. The paper reports heterogeneous and homogeneous experiments on Digit-5 and PACS, component ablations, communication/computation accounting, and a Digit-5 leave-one-domain-out evaluation.
Critical strengths
-
The method combines an explicit feature transformation with an architecture-independent communication format and does not require a public dataset or an auxiliary generator.
-
The reported Digit-5 improvement is substantial, and the paper goes beyond the main accuracy tables with three-seed results, homogeneous-model experiments, unseen-domain evaluation, and cost accounting.
Critical weaknesses
- Eqs. (5)-(7) apply one scalar RMS normalization and then average class means and diagonal variances coordinate by coordinate. Eqs. (9)-(13) subsequently use these coordinates to recenter and rescale another client's features. Equal dimensionality and equal RMS magnitude do not establish common semantic coordinates: a feature-channel permutation with a matching inverse permutation in the local classifier leaves that client's predictions unchanged, but changes the communicated statistics and hence the virtual features received by other clients. The stated training objective contains local and virtual-feature cross-entropy, but no explicit explanation or measurement showing how a consistent cross-model coordinate system emerges. This does not establish that the method cannot work empirically. It does leave unclear whether the transmitted variation represents domain structure, architecture-specific basis mismatch, or a mixture of both.
- Removing L_VFG from Eq. (16) leaves only local cross-entropy, with no parameter aggregation. Under matched initialization, data, and optimization, this should recover the standalone learning rule. Yet Table 3 reports 34.21%/30.57% for 'w/o VFG loss', while Tables 1-2 report 27.89%/26.85% for Standalone: gaps of 6.32 and 3.72 points. Moreover, the full method improves over mean-shift-only by 1.15 points on Digit-5 and 0.43 on PACS, and over global statistics by 1.24 and 0.53. These are much smaller than the headline baseline gains. The contribution of diagonal spread and domain-specific statistics therefore needs stronger isolation, particularly against a matched class-conditioned feature-statistic augmentation control using the same information budget.
Questions for authors
- What enforces or empirically demonstrates shared feature coordinates across the independently trained architectures?
- Why does 'w/o VFG loss' differ so substantially from Standalone? Please compare them with identical initial states, warmup, sample order, optimizer steps, and normalization-state handling, and describe any other remaining distinction.
- Can you provide matched controls for RMS calibration, class-conditioned moment augmentation, and the added diagonal-variance information, together with accuracy versus cumulative communication on both datasets? How were the adapted heterogeneous baselines tuned?