- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper introduces Socialized Continual Learning, in which previously socialized experts integrate newly arriving expertise while retaining specialization. Recursive Weight-Space Editing (RWSE) expresses integration for an editable linear mapping as a quadratic functional-alignment problem. It accumulates historical feature Gram matrices and derives a closed-form residual update, augmented by a nonlinear spectral penalty that protects directions with large historical eigenvalues. Experiments on CIFAR-100 and ImageNet-100 compare classification performance under 5- and 10-expert partitions, including a synthetic-feature setting and a regularization ablation.
Critical strengths
-
The quadratic formulation makes the trade-off between fitting incoming expert outputs and preserving the previous mapping explicit. Compressing historical feature information into a Gram matrix is a potentially useful way to avoid storing all prior feature batches.
-
The reported accuracy improvements span both datasets and both expert partitions, and the synthetic-feature experiments examine a relevant limitation of relying on original examples during collaboration.
Critical weaknesses
- There is a substantial gap between the multi-expert problem being claimed and the algorithm that is actually specified. Section 3 describes a collection of separately specialized experts that should all expand their capabilities, whereas Section 4 specifies one evolving shared linear mapping W. It is not clear whether a separate RWSE trajectory is maintained for every expert, how those trajectories differ, or how the updated collection realizes directionality for all experts. More fundamentally, Eq. (2) defines V_t = K_t W_t for the incoming expert, and the update then evaluates the same K_t through the previous shared mapping. This requires compatible input coordinates and output semantics at the edited layer. The paper does not adequately specify where these correspondences are extracted in independently trained ResNet-18 experts, how feature-coordinate mismatch is handled, how new class outputs are aligned, or which layers are edited.
- The evaluation does not reliably establish the claimed specialization, retention, and scalability benefits. First, the stated BWT formula is the mean of final minus initial task accuracy, for which larger values indicate better retention. However, Figure 4 labels BWT as lower-is-better, displays positive values, and the text treats the lowest BWT as best. This could be an omitted sign reversal, but the actual metric must be resolved before accepting the retention conclusion. Second, Table 2's Avg values appear to average expert and general accuracy equally: for RWSE on CIFAR-100, (69.06 + 65.43)/2 = 67.245, matching 67.25. If those columns cover 10 expert classes and all 90 remaining classes as the 10/10 description suggests, the class-balanced overall accuracy would instead be 65.793. If General is a single equally sized group or an intermediate stage, that protocol needs to be specified. The table therefore does not currently provide an unambiguous measure of the all-class versatility claim. These combined methodological and reporting gaps materially weaken the conclusions rather than amounting only to presentation issues.
Questions for authors
-
Please specify the complete ResNet-18 integration procedure: edited layers, sources of K_t and V_t, cross-expert coordinate and class alignment, handling of upstream feature changes, and the ordering of SGD and closed-form edits. Does every original expert retain a distinct updated model?
-
What quantity is actually plotted on the BWT axis? Please provide the signed per-task accuracy matrix and define the exact class groups, collaboration stage, and weighting used in Tables 1 and 2.
-
Can you reconcile the ablation arithmetic and report matched isotropic, linear-spectral, and nonlinear-spectral penalties with trial variability?
I have checked the discussion and found no response from the author. So my concerns remain unaddressed. I keep my original evaluation and rating.