TNLearn: An Open Source Python Package for Task-based Neurons
Meng Wang, Tieyun Li, Juntong Fan, Hanyu Pei, Jing-Xiao Liao, Yaodong Yang, Jianwei Ma, Fenglei Fan
cs.LG · cs.AI
Quick verdict
TNLearn: an open-source package for auto-designed task-specific neurons; valuable tool, but evidence rests on single runs with possible baseline under-tuning.
Summary
TNLearn is an open-source Python package (PyPI, GitHub, Read the Docs) that operationalizes 'task-based neurons': instead of the standard inner-product-plus-activation unit, each neuron's aggregation function is a symbolic formula automatically discovered from the task's data, then re-parameterized so its coefficients become learnable. The package offers four discovery engines — genetic programming (GPSymRegressor), LLM-guided search with an experience buffer (LLMSymRegressor), REINFORCE-based term selection with Ridge fitting (RLSymRegressor), and a differentiable tensor-decomposed regressor (PolyTensorRegressor) — plus scikit-learn-style MLPRegressor/MLPClassifier wrappers and drop-in PyTorch layers (e.g., TNConv2d). Experiments on seven OpenML tabular datasets show task-based neurons beating linear neurons (Table 3; e.g., bike R² 0.8838 vs 0.6702) and random-formula controls (Table 4; fried R² 0.9586 vs 0.1597), and outperforming four quadratic-neuron variants (Table 5). On MNIST, TNConv2d reaches 99.21% vs 99.05% for standard Conv2d after 5 epochs. Main limitations: Tables 3–5 report single point estimates without seeds, variance, or significance tests; the discovered-vs-random comparison changes architecture per row; three quadratic baselines collapse on fried (R²≈0.16), suggesting under-tuning; formula search uses at most 1,000 subsampled instances; and no runtime or compute cost is reported.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
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
- Networks built from task-based neurons (symbolic formulas discovered from task data) consistently outperform same-architecture linear-neuron networks across the seven OpenML benchmarks (Table 3).
- Formulas found by RLSymRegressor beat complexity-matched random formulas on all seven datasets with fewer parameters, showing symbolic regression extracts real task priors rather than coincidences (Table 4).
- Task-based neurons achieve superior or competitive performance against four published quadratic-neuron designs when all are expressed inside TNLearn (Table 5).
- On MNIST, TNConv2d converges faster early and finishes at 99.21% versus 99.05% for standard Conv2d over five epochs (Table 6).
- TNLearn is the first systematic, user-friendly open-source materialization of the task-based-neuron concept (Sections 1–2).
Strengths
Four genuinely distinct discovery engines behind one contract
Section 2.1 and Appendix C implement GP with a custom inner-product operator (C.1), an LLM searcher with a multi-island experience buffer and BFGS coefficient fitting (C.2), REINFORCE term selection with Ridge fitting (C.3), and a differentiable gated polynomial/CP-interaction/periodic regressor (C.4, Eq. 4) — all emitting the same homogeneous aggregation form, so users can swap engines without rewiring.
Evaluation includes controls most library papers omit
Table 4 compares discovered formulas against complexity-matched random formulas, and Table 5 re-expresses four published quadratic-neuron designs (Bu & Karpatne 2021; Fan 2018; Goyal 2020; Xu 2022) inside the same framework for an apples-to-apples setting.
Strong reproducibility infrastructure
Hedged, mostly non-hyperbolic claims about method scope
Practical integration surface
Weaknesses
No uncertainty quantification in the core evidence
Likely unfair quadratic baselines in Table 5
Architecture confounds in the key controls
Thin data slice for the central 'task prior' premise
No cost accounting despite efficiency claims
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
TNLearn (arXiv 2609.27564, cs.LG) is an open-source software paper that packages the 'task-based neuron' concept from the authors' companion line of work (Fan et al., TPAMI 2026): a neuron's aggregation function is a symbolic formula — sums of powers, inner products, trigonometric terms — discovered automatically from task data by symbolic regression, then re-parameterized so its constants become learnable weights. The paper makes four claim types: (1) conceptual — neuron design is a design axis complementary to architecture ('dimension augmentation', Section 1); (2) engineering — TNLearn automates neuron and network construction with four search engines; (3) empirical — task-based neurons beat linear neurons, random formulas, and four quadratic-neuron variants (Tables 3–5); (4) dissemination — PyPI/GitHub/Read the Docs availability and 'PyTorch ecosystem project' status, the latter unverifiable from the text.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 a task-based neuron?
A neuron whose aggregation function is not the standard weighted sum but a symbolic formula automatically discovered from the task's training data (for example, sums of powers, inner products, or trigonometric terms). After discovery, the formula's numeric constants become learnable parameters before the neuron is placed inside a network.
Is TNLearn free, and how do I install it?
Yes, it is open source on GitHub (NewT123-WM/tnlearn) and installable from PyPI with 'pip install tnlearn'; a specific version can be pinned (e.g., 0.1.0). It requires Python 3.8+ and builds on PyTorch, scikit-learn, and NumPy. Documentation with API reference and examples is at tnlearn-documentation.readthedocs.io.
How much better are task-based neurons than regular neurons?
In the paper's tables they beat linear neurons on all seven OpenML datasets (largest gap: bike-sharing R² 0.8838 vs 0.6702; smallest: adult accuracy 0.8582 vs 0.8546) and random formulas on all seven. MNIST gains are modest (99.21% vs 99.05% after five epochs). All tabular results are single-run point estimates, so treat the magnitudes cautiously.
How does TNLearn differ from quadratic-neuron libraries like QuadraLib?
QuadraLib (Xu et al., 2022) optimizes one fixed neuron family, the quadratic neuron. TNLearn instead discovers a dataset-specific formula per task using four engines (genetic programming, LLM search, reinforcement learning, differentiable tensor decomposition), and can itself express the quadratic variants as baselines — which it reports matching or beating on five benchmark datasets.
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 1–6 are referenced but only captions and Table 6 numbers survived extraction; the architecture diagram and accuracy curves could not be inspected.
- Some table content is typographically mangled in the extracted text (notably the satimage formula in Table 2 and formula notation in Table 4); individual symbols may be extraction artifacts rather than paper errors.
- Claims about the codebase (rigorous unit testing, benchmark directory) and 'PyTorch ecosystem project' status cannot be verified from the PDF text alone.
- The paper is a v1 preprint (23 Sep 2026); peer-review status is unknown.
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.