Interpreting the predictions of neural network classification based on a Taylor Coefficient Analysis (TCA)
Markus Klute, Artur Monsch, Lars Sowa, Roger Wolf
cs.AI · physics.data-an
Quick verdict
TCA formalizes second-order Taylor feature attribution for NN classifiers, shown on toy and LHC-like tasks; strong pedagogy, but no XAI baseline comparison.
Summary
The paper introduces the Taylor Coefficient Analysis (TCA), a taxonomy in which every non-empty subset of input features corresponds to a Taylor coefficient of a neural network's output ŷ expanded in the inputs: first-order coefficients capture gradient-like saliency, second-order ones capture curvatures and pairwise correlations. Coefficients are aggregated as the mean absolute value ⟨t⟩ over samples and can be tracked epoch by epoch. Two Gaussian toy tasks, with 100-train ensembles and 68% CIs (Table 1), illustrate the mechanics: in the correlated-covariance example ⟨t_x1x2⟩ overtakes first-order features after roughly 700 epochs, mirroring a three-phase learning curve. An LHC-style task (ggH→ττ vs tt̄, MadGraph+Pythia+DELPHES, 20 input features, AUC 0.99) ranks Δφ(pTmiss, pTτℓ) and its self-term as most influential; mass variables rank low because ŷ saturates there, and the deliberately spurious φℓ lands 16th of 20 as expected. Practical sections add a cumulative-count summary statistic, the caveat that ReLU networks carry no hidden-layer information in higher-order coefficients, and a Kolmogorov–Smirnov study on one model stopped at AUC 0.95 supporting second-order truncation. Limitations: validation relies on the authors' physics intuition rather than head-to-head comparison with SHAP/LOO/gradient baselines, and the second-order-sufficiency claim is extrapolated from a single task.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Every non-empty subset of input features defines a generalized feature via its Taylor coefficient, and Taylor-series convergence makes this taxonomy of influences on ŷ complete (Definition 1, Section 2).
- TCA reveals which input properties drive a prediction and tracks when during training each is learned, demonstrated on two Gaussian toys and an LHC-style ggH→ττ vs tt̄ classification task.
- For a task representative of typical LHC complexity, a Kolmogorov–Smirnov study (Section 4.4.3) supports truncating the TCA at second order.
- For ReLU networks, higher-order Taylor coefficients carry no hidden-layer information beyond first order — an intrinsic property of the architecture, not a flaw of TCA (Section 4.4.2).
Strengths
Clean formal taxonomy generalizing feature importance
Definition 1 (Section 2) maps every non-empty subset α of inputs to a Taylor coefficient t_α, unifying saliency (first order), curvatures t_xjxj and pairwise correlations t_xixj (second order); Definitions 2-3 formalize decision plane and boundary used consistently in Figs. 3, 4, 8.
Quantitatively careful toy studies
Table 1 reports ⟨t_α⟩ after 120 and 1,000 epochs with 68% CIs from 100 random-initialization trainings; Example-2's ⟨t_x1x2⟩ grows from 10.40×10^-2 to 18.06×10^-2 while Example-1's falls to 7.23×10^-2 — a well-uncertained demonstration that TCA tracks correlation learning (Section 3.2).
Built-in sanity check in the LHC application
Honest ReLU limitation analysis
Released code foundation
Weaknesses
No empirical baseline comparison
Circular validation
Second-order sufficiency rests on one model
Aggregation under-ranks saturated but decisive features
Reproducibility gaps for the LHC study
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a method paper from a KIT particle-physics group formalizing the Taylor Coefficient Analysis (TCA), previously applied informally in CMS-adjacent analyses (Refs. [43,44,36,39,11]). The central conceptual claim is Definition 1 (Section 2): every non-empty subset α of the n input observables maps to a Taylor coefficient t_α of the expansion of the network output ŷ, so gradients, curvatures t_xjxj, and pairwise correlations t_xixj become 'generalized features' of order κ = |α|; since the Taylor series converges to ŷ, the taxonomy is claimed complete. A second claim is practical: for tasks of typical LHC complexity, truncation at second order suffices (Section 4.4.3). Support comes from two Gaussian toys (Section 3), an LHC-style ggH→ττ vs tt̄ task (Section 4), and guidance on strategy, a cumulative-count summary, and the ReLU caveat.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Taylor Coefficient Analysis (TCA) in neural network interpretability?
TCA expands a neural network's output ŷ as a Taylor series in its input features. Each Taylor coefficient — first order for single features (gradient-like sensitivity), second order for curvatures and feature pairs — is treated as a 'generalized feature', and the average magnitude of these coefficients over a data sample quantifies how much each influences the prediction. It was developed for particle-physics event classifiers.
Does TCA work with ReLU networks?
Only partially. Because ReLU's second derivative is zero, all higher-order Taylor coefficients probe just the output layer and carry no beyond-first-order information about hidden layers. The paper therefore uses tanh activations for full second-order analysis and notes smooth activations are needed when second-order TCA is the goal.
How many Taylor orders are needed for LHC-like classification tasks?
The paper's Kolmogorov–Smirnov study on one 20-feature model found second-order Taylor approximations statistically consistent with the true output up to input shifts of about 1.4 standardized units (p = 0.056 at 1.8), suggesting second order usually suffices for tasks of typical LHC complexity — though this was tested on a single model.
Is TCA code available?
Yes. The authors provide the PyTorch 'TaylorAnalysis' package on Zenodo (DOI 10.5281/zenodo.14998996), supporting efficient, multithreaded computation of the Taylor coefficients.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-10 appear only as captions, axis labels, and prose descriptions in the extraction; all figure-level judgments rely on the paper's own textual descriptions.
- The extracted text ends with the complete reference list after the Summary section; no truncation was detected, so no content gaps are claimed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.