{"id":"9e6b59f7-1cbf-4e84-9ebc-b35fe5feb937","arxiv_id":"2607.29621","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"CENDRe recovers the time- and frequency-domain patterns that drive a trained time-series CNN's decisions, choosing the number of concepts automatically via silhouette-guided clustering.","lead":"CENDRe is a new explainability tool that shows which recurring patterns — in both the time and frequency domains — a trained time-series CNN uses to make decisions, without requiring a user-chosen number of concepts. It matters because engineers inspecting fault-detection and monitoring models can check whether predictions hinge on physically meaningful frequency bands rather than on artifacts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claims depend on the assumption that trained CNNs rely on the injected synthetic primitives; the paper itself flags this as strained and observes shortcut learning, so sRC/sIC and frequency-band recovery may be measured against a false reference.","rationale":"The reader's weakest_assumption is exactly the premise I identify: the synthetic evaluation assumes model behavior is strongly predictable from dataset construction, and the paper itself observes that this property is strained. My reading of Sec. 4, Sec. 5, and App. B.2 confirms that this is the single most load-bearing assumption for the central empirical claims. It affects both time-domain metrics (sRC, sIC) and the frequency-domain recovery claim, and it carries over to the qualitative real-data interpretation, since no independent faithfulness check is provided there. The paper has real independent strengths — the Wirtinger derivation in App. A is careful, the silhouette-based K selection is evaluated with a quantitative sensitivity analysis, and the three-architecture/eleven-seed protocol is thorough. These strengths do not, however, remove the need to verify that the models actually rely on the injected primitives. The proposed ablation test is a direct, low-cost way to settle whether the concern lands: if the trained models are insensitive to removal of the annotated primitives, the quantitative headline claims lose their reference; if they are highly sensitive, the current sRC/sIC numbers can be trusted. Since the paper already conditions its conclusions on this premise being acceptable, I do not move the verdict; I keep the reader's CONDITIONAL position, with the condition being that the reliance premise be empirically verified.","tokens_in":38925,"tokens_out":6847,"duration_ms":73252,"concrete_test":"Counterfactual ablation on the paper's trained models: for each SYNTHETICLOCAL validation window, replace the annotated primitive support with the class-agnostic noisy baseline; for each SYNTHETICFREQUENCY window, zero the injected band bins in the rFFT and invert. Measure the drop in the true-class logit/accuracy. Compare with the same drop when the top 10% of each CENDRe mask's support is ablated. If primitive ablation is not substantially more damaging than abating equal-area control regions, the annotated masks are not the model's decision regions and the sRC/sIC/band-recovery claims lack a valid ground truth. If primitive ablation is clearly decisive, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claims — representation correctness comparable to state-of-the-art, significantly higher importance correctness, and frequency masks that recover ground-truth bands — all rest on the premise that models trained on the synthetic datasets use the annotated primitives as their decision regions. The paper explicitly admits this premise is strained: Sec. 4 states the assumption, Sec. 5 reports that the model exploits only a portion of a primitive (a shortcut-learning pattern), and App. B.2 says the synthetic benchmarks are only informative when model behavior is strongly predictable from the dataset construction, a property that is already strained. If the model uses only fragments of a shape, or time-domain artifacts of injected spectral peaks rather than the frequency bands themselves, then the binary primitive masks are the wrong reference. Under that mismatch, sRC penalizes CENDRe's focused masks and rewards ECLAD-ts's broad masks; sIC measures alignment with a phantom ground truth; and the SYNTHETICFREQUENCY band-recovery claim is no longer evidence about what the model actually computes. Figure 1 itself is consistent with this risk: CENDRe weights square corners more strongly while ECLAD-ts paints the full square, which is what one would expect if the CNN relies on edge-like shortcuts. The real-data frequency claims then lack an independent faithfulness check. This is not a disagreement with consensus; it is a correctness risk in the evaluation reference, and the paper's own admissions make it the most load-bearing point to resolve.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"CENDRe is a post-hoc concept extraction method for 1D CNNs on time series. It builds per-timestep LADs from probe layers, clusters them in two stages (micro-clusters then concepts), selects the number of concepts automatically via silhouette maximization, localizes concepts with gradient masks in the time and frequency domains through virtual inspection layers, and computes per-concept class importance scores. The paper claims three advances: the first CE method to extract concepts in both time and frequency domains, automatic concept-number selection, and gradient-based localization that concentrates on regions the model actually uses. Experiments on synthetic datasets (SYNTHETICLOCAL, SYNTHETICFREQUENCY) and real bearing-fault data (CWRU, BearingPD) are reported, with the headline claims that CENDRe achieves representation correctness comparable to prior CE methods and significantly higher importance correctness, and that its frequency masks recover ground-truth bands on synthetic data and characteristic fault frequencies on CWRU.","tokens_in":39136,"tokens_out":4209,"duration_ms":46221,"significance":"If the empirical claims hold, CENDRe would be a useful step forward: it is the first time-series CE method with a principled frequency-domain concept visualization, the silhouette-based K selection replaces a manual sweep, and the derivation of the spectral mask (App. A) is clean and parameter-free. The experimental investment is unusually thorough: three architectures, 11 seeds, Holm-corrected Wilcoxon tests, sensitivity analysis over the micro-cluster count J, and disclosed HDBSCAN failure rates. These strengths are real. However, the central evaluation premise — that models trained on the synthetic data rely on the injected primitives — is explicitly admitted in the paper to be strained, and the comparison protocol contains an oracle-K bias. These issues affect the load-bearing claims about importance correctness and frequency-band recovery, so the headline result is not yet established at the level the abstract implies.","major_comments":[{"comment":"The synthetic benchmark premise is load-bearing and is explicitly admitted to be strained. The paper states that CE evaluation 'assumes a model trained on these datasets relies on them' (Sec. 4), then reports a 'shortcut-learning pattern' where models exploit only a portion of a primitive (Sec. 5), and App. B.2 says the benchmarks are informative only if model behavior is strongly predictable from dataset construction, 'a property that is already strained.' If the model uses fragments or time-domain artifacts of injected spectral peaks, then the binary primitive masks are the wrong reference, and sRC/sIC measure alignment against a phantom ground truth. The abstract's 'significantly higher importance correctness' and the frequency-band recovery claims rest on this reference. Please add a direct check of model reliance on the primitives — e.g., accuracy drop when the primitive regions are","section":"Sec. 4 and App. B.2"},{"comment":"The comparison protocol is biased by an oracle K sweep for fixed-K methods. ECLAD-ts, MultiVISION, and CENDRekMeans are evaluated at the K in {2,4,6,8,10} with the highest sRC, whereas CENDResilhouette/HDBSCAN run once with automatic K. Thus the sRC gap on SYNTHETICLOCAL (Tables 5: ECLAD-ts 0.741 vs. CENDResilhouette 0.651) reflects partly selection on the metric. The paper acknowledges this (Sec. 5), but it is load-bearing because the 'comparable representation correctness' and 'higher importance correctness' claims are relative to an oracle-tuned baseline. Please report fixed-K methods at a non-oracle K (e.g., K=4, median of the sweep) or average over the sweep, and re-test the pairwise significance in Fig. 9 under that protocol.","section":"Sec. 4.2 and App. D.1.4"},{"comment":"Only CENDRe variants are evaluated on SYNTHETICFREQUENCY because the baselines lack a frequency-domain counterpart. The claim that frequency masks 'recover the ground-truth bands' is therefore not established against any baseline, and it inherits the unverified premise of the first major comment. The time-domain masks for this family are reported as uninterpretable (Fig. 3), so the only evidence of spectral grounding is the frequency-domain mask itself. Without showing that the model's predictions depend on those bands — e.g., by spectral masking of the injected bands and measuring accuracy drop, or by training a model on a version with the bands removed — the correspondence could be an artifact of the gradient propagation rather than faithful explanation. Please add a faithfulness check for the frequency domain.","section":"App. B.2 / Sec. 5 (SYNTHETICFREQUENCY)"}],"minor_comments":[{"comment":"The header 'CENDREACsilhouette' should be 'CENDRe silhouette' to match the text and other figures.","section":"Fig. 9"},{"comment":"The factor 0.5 in the definition of Ik,ch is unexplained; the contrast term is already a difference of means, so the constant scaling appears to carry no information. Please clarify or remove.","section":"Eq. (5)"},{"comment":"For the DWT visualization, the specific wavelet family and number of decomposition levels are not stated. Since the authors require exact invertibility, these details are needed for reproducibility.","section":"App. D.1.2"},{"comment":"No code or data availability statement is provided. Given the complexity of the pipeline (virtual inspection layers, clustering, metrics), a public implementation would substantially improve reproducibility.","section":"General"},{"comment":"The CE method is fit on only 256 samples. With J=50 micro-clusters, this is about 5 samples per micro-cluster before aggregation; a brief comment on how this interacts with silhouette stability would be useful.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about the strain in its synthetic evaluation, which is commendable, but that strain is central to the headline claims. The method itself and the spectral derivation appear sound, and the experimental infrastructure is above the field average. I would be willing to accept after the authors add a direct verification that trained models actually use the annotated primitives (or reformulate the metrics accordingly) and after removing the oracle-K bias from the comparison. No concerns about novelty or scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a serious look. The core new capability — global concept explanations in the frequency domain for time-series CNNs — is real and, as far as I can tell, genuinely new. The method combines the authors' LADs, virtual inspection layers, and a gradient presence-score into a coherent pipeline, and the Appendix A spectral mask derivation is clean, parameter-free Wirtinger calculus. The empirical apparatus is a cut above the usual XAI paper: three architectures, 11 seeds, Holm-corrected Wilcoxon tests, a quantitative sensitivity analysis for J, explicit HDBSCAN failure rates, and honest runtime accounting. The paper is also transparent about its own limitations, which counts for something.\n\nThe soft spots are real but proportionate. The biggest one is the one the stress test flags: the synthetic evaluation assumes trained CNNs rely on the injected primitives, and the paper itself admits that premise is strained and that models exploit only fragments of primitives (shortcut learning). If the model actually uses corners rather than the full square, then the ground-truth primitive mask is the wrong reference. Under that mismatch, sRC penalizes CENDRe's focused masks and rewards ECLAD-ts's broad ones — exactly what the numbers show. So the headline 'significantly higher importance correctness' is measured against a reference that may not correspond to what the model computes. The frequency-band recovery on synthetic data is externally anchored, but the same strain applies there: if the model keys on time-domain artifacts of the injected peaks, band recovery is not evidence about the computation.\n\nTwo smaller issues. First, sIC is partially self-referential: the importance term in Eq. 7 is derived from the same masks the method produces. It is not fully circular, because alignment to primitives is also involved, but it should be checked against an independent importance benchmark. Second, baselines get an oracle-K sweep while automatic variants run once, and CENDRe trails ECLAD-ts by about 0.09 sRC with p<0.01 on every SYNTHETICLOCAL dataset. Calling that 'comparable' in the abstract is too generous, even if the explanation (shortcut learning plus oracle bias) is plausible.\n\nThe real-data claims are qualitative, with no quantitative comparison to spectral-attribution baselines like virtual inspection layers, which are the closest competing tool. I would not call any of this fatal. The method is sound, the derivation is solid, and the honesty is refreshing. The flaws are in the evaluation framing and the abstract's wording, not in the core idea.\n\nSend it to peer review. It deserves a serious referee, and the authors have shown they can respond to careful criticism. I would cite it if I worked on time-series interpretability.","headline":"Genuinely new frequency-domain concept extraction with unusually careful experiments, but the synthetic ground truth is strained and the abstract overstates the sRC comparison.","tokens_in":39787,"tokens_out":2088,"would_cite":true,"duration_ms":25605,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CENDRe introduces concept extraction for time-series CNNs that works in both the time and frequency domains, selects the number of concepts automatically, and reports higher importance correctness than prior baselines on synthetic benchmark","keywords":["concept extraction","time-series classification","frequency-domain explanations","CNN interpretability","silhouette clustering","virtual inspection layers","bearing fault diagnosis","gradient-based localization"],"falsifier":"Train a CNN on a synthetic square-triangle dataset, then replace every square and triangle with the noisy baseline and measure accuracy: if accuracy stays high, the model does not rely on the annotated primitives, and the correctness scores against those masks would be high while explaining nothing real. For the frequency claim, verify the model uses only a single injected band (for example, by ablating that band and confirming accuracy drops), then check whether the aggregate frequency mask concentrates in that band; if the mask mass sits elsewhere, the virtual-inspection gradient has failed","tokens_in":38659,"feed_emoji":"⚙️","tokens_out":4749,"duration_ms":57527,"temperature":0.7,"pith_summary":"The paper introduces CENDRe, a concept extraction method for convolutional time-series classifiers that produces global explanations in both the time and frequency domains. It claims to be the first such method to visualize concepts in the frequency domain, and it removes the need to preselect the number of concepts by deriving that number from the latent structure of the data. On synthetic benchmarks with injected ground-truth primitives, CENDRe matches prior methods on representation correctness while achieving significantly higher importance correctness. On real bearing-fault signals, the extracted frequency masks reportedly include the characteristic fault frequencies that experts inspect, offering evidence that time-domain-only explanations cannot provide. If correct, CENDRe gives practitioners a way to see which spectral features a time-series CNN actually uses and how strongly each concept drives each class.","feed_headline":"CNN concepts now explainable in frequency domain","feed_subtitle":"Method recovers the spectral bands driving predictions and sets concept count automatically.","key_machinery":"The load-bearing object is the per-timestep Local Aggregated Descriptor (LAD): a vector formed by concatenating upsampled activations from several probe layers at each timestep, so that similar local patterns produce similar descriptors regardless of position. Concept discovery runs in two stages — mini-batch k-means compresses LADs into micro-centroids, then silhouette-guided hierarchical aggregation groups these into concepts, determining K automatically. The transfer mechanism is the virtual inspection layer: a fixed, differentiable, invertible transform (for example, the real-input fast Fourier transform) inserted into the computation graph so that concept-presence gradients can be evalu","core_discovery":"CENDRe claims that a single concept-extraction pipeline can yield concept masks in any invertible, interpretable signal domain. It discovers concepts by clustering per-timestep latent descriptors in two stages, using silhouette-guided aggregation to set the number of concepts K automatically. Each concept is then localized by backpropagating gradients of a contrastive presence score — a measure of how much a latent representation resembles its assigned prototype rather than other concepts' prototypes. Because these gradients are propagated through a differentiable invertible transform such as the Fourier transform, the same concept can be visualized in the frequency domain. On synthetic data","pith_inferences":["Editorial inference: if silhouette-selected K remains stable across seeds and datasets, the method could be used as a cheap model audit before deployment — flagging concepts that align with spurious spectral bands rather than physically meaningful ones; the paper reports stability but does not make this application claim.","Editorial inference: the same virtual-inspection gradient could serve as a spectral shortcut detector: compare the aggregate frequency mask against known physical fault bands and flag any large mass outside them; this test would extend the method's diagnostic use beyond what the paper explicitly evaluates.","Editorial inference: because LADs rely on approximate translation equivariance, CENDRe's concept discovery is likely to degrade on architectures without that property, such as transformers or state-space models; a position-aware or attention-weighted variant would be a natural next step, which the paper mentions only as future work."],"forward_implications":["Frequency-domain concept explanations become available for time-series CNNs without retraining or modifying the model, enabling inspection of spectral features that time-domain methods miss.","The number of concepts is selected from latent structure rather than set by the user, removing a bias-inducing hyperparameter and making concept extraction less dependent on manual sweeps.","Gradient-based localization concentrates masks on the regions driving each concept, which the paper ties to higher importance correctness on synthetic benchmarks.","The same virtual-inspection mechanism extends to other exactly invertible transforms, such as short-time Fourier and wavelet transforms, giving time-frequency views of the same concepts.","On bearing-fault signals, concepts localize in frequency bands that experts use for diagnosis, supporting model verification in a safety-relevant application."],"fun_headline_variants":["CENDRe explains CNN predictions in time and frequency","Time and frequency concepts auto-extracted from CNNs","CENDRe: frequency-domain concept masks for CNNs","Frequency-domain concepts automatically discovered in CNNs","CENDRe sets concept count and shows frequency patterns"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The evaluation assumes that a CNN trained on the synthetic datasets actually relies on the injected class-discriminative primitives; the paper itself acknowledges that models often exploit only a portion of a primitive (a shortcut-learning pattern) and that this assumption is already strained, so if the models chiefly use other cues, the reported correctness scores measure alignment against the wrong reference.","fun_headline_variants_meta":{"raw":{"variants":["CENDRe explains CNN predictions in time and frequency","Time and frequency concepts auto-extracted from CNNs","CENDRe: frequency-domain concept masks for CNNs","Frequency-domain concepts automatically discovered in CNNs","CENDRe sets concept count and shows frequency patterns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001102,"raw_usage":{"total_tokens":4438,"prompt_tokens":756,"completion_tokens":3682,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":3621}},"tokens_in":500,"tokens_out":3682,"duration_ms":24766,"temperature":1.0,"reasoning_tokens":3621,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T03:20:05.651404+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a CNN on a synthetic square-triangle dataset, then replace every square and triangle with the noisy baseline and measure accuracy: if accuracy stays high, the model does not rely on the annotated primitives, and the correctness scores against those masks would be high while explaining nothing real. For the frequency claim, verify the model uses only a single injected band (for example, by ablating that band and confirming accuracy drops), then check whether the aggregate frequency mask concentrates in that band; if the mask mass sits elsewhere, the virtual-inspection gradient has failed","supporting_citations":[],"review_version":1}