- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper points out that pointwise training data attribution misses the information carried by the relations between training data. For instance, two data points with the same pointwise attribution score can exhibit different impact on other data, and this impact could even be asymmetric. To represent both pointwise and relational information, the paper proposes Relationship-Aware Training Data Attribution, which is a graph-based attribution method. Since processing all relations between data has quadratic complexity in the number of data points, the paper instead addresses the scores of data groups. The node of the graph represents the attribution score of the group, and the directional edge of the graph represents the extent of the difference in the target group score when the source group is progressively removed.
Critical strengths
The most interesting point I found is the asymmetry of the relationship between the groups. To the best of my knowledge, current training data attribution literature that addresses the relation between data treats it as symmetric, so if the paper’s claim is true, it would be a different type of relationship compared to current methods. Related to this, I’d like to see whether substituting the in Theorem 2 with the existing training data attribution methods such as TRAK or influence functions still yields the asymmetric relation or not.
Critical weaknesses
- The asymmetric result in Figure 2 could be an artifact from the randomness of retraining. Figure 2 measures how much the TRAK score of the target data changed when removing the source data and then retraining the model, which depends on training randomness such as mini-batch selection. This randomness could not be controlled by fixing the training seed since different data removals affect the mini-batch selection even with the same seed. So the asymmetric result in Figure 2 couldn’t be evidence of the relation’s asymmetry. The same problem arises in the paper’s method, which defines the edge as the effect of retraining on TRAK ranking. Without an analysis on retraining noise, it is hard to know whether the proposed edge reflects the actual relation between training groups.
- The paper uses the change of TRAK score as the main metric for analysis. However, TRAK itself is an approximation of the retraining effect. So the current analysis could not identify whether the results come from the inherent error of TRAK or the relation is actually important. One of the possible alternative metrics is leave-group-out retraining, which is a standard metric of training data attribution; however, I expect it to yield a symmetric relation. Then is the asymmetric relation derived by the paper an artifact from the inherent error of TRAK?
- The proposed Relationship-aware training data attribution needs repeated model retraining to get edges, which raises concern about its prohibitive cost. For large-scale models or datasets, it could be challenging to apply the method due to this constraint.
Questions for authors
- Could the authors distinguish the effects of the randomness of retraining and of the relation between training data in Figure 2? How much of the whole effect does the randomness effect account for? And is the relation still asymmetric after removing the randomness effect?
- Could the authors explain the reason for using TRAK as the metric throughout the paper? If the leave-group-out retraining effect is used instead of TRAK, is the relation still asymmetric?
- Could the proposed method extend to large-scale models or datasets? Could the authors provide the wall-time of the experiments of the paper?
From AC.
Dear Authors,
Given the deadline on 18 November, please submit your responses to the reviews to facilitate a productive discussion with the reviewers.
Thanks, AC