By Stefano Colamonaco
Review Details
Reviewer has chosen not to be Anonymous
Overall Impression: Good
Content:
Technical Quality of the paper: Average
Originality of the paper: Yes, but limited
Adequacy of the bibliography: Yes, but see detailed comments
Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Limited
Organization of the paper: Needs improvement
Level of English: Satisfactory
Overall presentation: Average
Detailed Comments:
Recommendation: Major revision
Summary: The paper designates the encoder layer of a time series transformer as a concept bottleneck layer. Its components are pushed during training to be similar to two predefined concepts, a linear AR surrogate forecaster and other features, using a CKA term in the loss. The residual connection around the layer is removed. The method is applied to Autoformer on one synthetic and six benchmark datasets, and tested with an activation patching experiment on shifted timestamps. The setting and results are interesting but I think the paper has a lot of room for improvement.
Reasons to Accept:
- Interpretability by design for time series is under-explored, and the idea of encouraging concept locality instead of searching for it post hoc it is a direction that fits with the journal's topics.
- The method is a regulariser, so it is applicable to existing architectures easily.
- The patching experiment is interesting. Most CBM papers stop at correlational evidence.
- The component plots (Fig.8) are a good support to the paper's theses and show that the method is valid.
- the results appear genuine and correctly report the strong baselines.
Reasons to reject (and room for improvement):
- Framing
The introduction sets up neuro-symbolic AI and mechanistic interpretability as converging fields, then drops both. Neither does any work in the rest of the paper, which is a CBM paper.
The claim that CBMs are not usually considered neuro-symbolic (page 4) it is a strong position that does not find correspondence in the bibliography cited by the authors: Barbiero et al. (2023) is titled Interpretable Neural-Symbolic Concept Reasoning. Poeta et al. is a survey of concept-based XAI, not an authority on what counts as NeSy.
More basically, it is unclear what is symbolic in the method. The concepts are hidden activations of a linear model and a scalar calendar feature. No symbolic operation is performed on them. The authors should either defend the NeSy framing with actual content or drop it and position the paper as concept-based interpretability for forecasting.
- Structure
The Background and Related Work section needs restructuring. The CKA definition and the transformer formalism are background material and would sit more naturally in the Method section. The remainder is organised around the components the authors build on rather than around the surrounding literature. I would suggest reorganising it thematically: concept bottleneck models and their known failure modes, interpretability for time series, and representational similarity measures used as training signals (I would also consider bringing more literature on this area).
The claim that "a line of work on CBMs for time series was initiated by our earlier preprint" understates prior contributions, given Ferfoglia et al. (2024). A formulation such as "concurrent and subsequent work includes..." would be more accurate. The authors should also state explicitly what this manuscript adds relative to the earlier preprint.
- What interpretability means here
The paper never defines interpretability, and what it proposes is weaker than the general notion.
CKA is a batch-level similarity measure, invariant to scaling and orthogonal transformation. It does not follow that you can read the concept value off that component for a given input, or that the component encodes only that concept. So the usual CBM affordances are gone: no per-instance concept explanation, no direct intervention setting the concept by hand.
What the reviewer thinks the method does give is concept localisation: the influence of a known concept concentrates in a known component, which makes a targeted edit possible. That is still useful.
There is also no interpretability measure independent of the loss. CKA is optimised and then reported as the evidence. Some literature shows that CKA can be driven high by manipulations that do not reflect real correspondence (e.g. Correcting Biased Centered Kernel Alignment), which matters more when you optimise it directly. Adding a probe would make the claims grounded and stronger.
- The intervention
This is activation patching, not a classic CBM intervention. Activations come from a second pass of the same model, not from a human-supplied concept value. The method section mostly gets this right, but would be necessary to separate the two notions and adjust the claims.
The experiment is run on electricity only and I would specify it in contribution 3. Or run the experiment also on other datasets.
- The bottleneck is not a bottleneck
This is the leakage that is well known in CBM literature. A component can score high and still carry unrelated information.
I won't consider this one of the method's major limitations because, despite its well-known nature, there are still studies working on it and others that simply acknowledge the problem. For this paper, I would follow the second approach.
See also Mahinpei et al. (2021).
- Figures
Fig. 1 is the most important and the least clear. The scissors is unexplained, the orange and white slabs are undefined, there is no legend, and the reader cannot reconstruct the method from it. Redraw with labelled components and an explicit marker for the cut residual, matching Fig. 2's vocabulary. An example of labels would also be useful there.
Fig. 7a lacks the ± values that 7b has.
- Results: apparent errors in the tables
The Traffic FF-bottleneck MSE of 0.393 is potentially wrong (?). It reaches MAE of 0.377 against a baseline of 0.619 MSE and 0.387 MAE, so a 36% drop in MSE accompanies a 2% drop in MAE.
The "No bottleneck" column differs between Tables 3 and 4 and I am not sure this is correct. That model has no concepts and no free component, so c=3 and c=2 should not apply to it. It still differs on all six datasets. I think this requires an explanation.
- Results: the comparison does not seem fair
I have a concern about the comparison in Table 2 between the bottleneck models and the Wu et al. column. The means in that column match the published Autoformer figures to three decimals on all six datasets, so they appear to be quoted from the original paper and not re-run; yet the setup described on p. 7 differs from it, with three encoder layers against Autoformer's two. The gap to Wu et al. therefore seems to mix the effect of the bottleneck with an encoder-depth change and a cross-paper comparison.
- Results: the two variants diverge and this is not discussed
FF outperforms Att on five of six datasets, and the intervention succeeds cleanly only for FF. Figure 10a shows the Att intervention performing worse than no intervention for shifts of roughly 5 to 13 hours, a negative result the text passes over. No reason is offered of why splitting concepts across attention heads should work less well than slicing the feed-forward output.
- Underspecified details
Several details needed to reproduce or assess the method are missing. Equation 1 gives the CKA formula but the text never says whether the features are centred, and it does not explain how a similarity score is computed between a multi-dimensional model component and the single hour-of-day feature.
The AR surrogate is also underspecified: the input length is never given for the real datasets, and it is not stated whether the model is fit per variate or jointly, or which of its quantities serve as the concept representation.
Finally, the model configuration should be reported in full so that "an original Autoformer of similar size" can be checked against the bottleneck models, and I would encourage the authors to release their code, since the contribution is a training framework whose value depends on others being able to apply it.
Reasons for the recommendation
I don't think the paper is ready for publication in its current state, but there is room for improvement. With corrected framing (the most important part for me), verified tables and experiments, a probe-based interpretability measure, I would expect to recommend acceptance.
