ChainzRule: Sample-Efficient, Robust Deep Learning Across Tabular, NLP, and Vision Tasks

# Sample-Efficient, Robust Deep Learning Across Tabular, NLP, and Vision Tasks
Source: [https://arxiv.org/html/2605.24340](https://arxiv.org/html/2605.24340)
###### Abstract

Production deep learning systems across enterprise domains operate under constraints that academic benchmarks routinely obscure: labeled data is expensive, inference budgets are tight, and models that cannot explain their behavior are difficult to trust and maintain\. We presentChainzRule \(CR\), a neural architecture replacing typical activations with learnable polynomial layers governed byDifferential Regularization \(DREG\), a layer\-wise Jacobian penalty computed analytically during the forward pass at standard inference cost\. The core claim is that bounding intermediate derivatives forces the network toward low\-frequency, structurally stable representations, simultaneously reducing dependence on labeled data volume, improving robustness to distribution shift, and providing a measurable, gradient\-based handle on model behavior\. Evaluated across five domains, CR achieves85\.71%±2\.01%85\.71\\%~\\pm~2\.01\\%on Pima Diabetes \(statistically superior to SVM and XGBoost\),46\.20%±0\.37%46\.20\\%~\\pm~0\.37\\%on SST\-5 sentiment classification with a frozen encoder \(superior to RNTN using approximately 5% of its training data\),55\.79%55\.79\\%on SST\-5 with a fine\-tuned BERT backbone \(versus BERT\-base linear head at54\.9%54\.9\\%\),70\.17%70\.17\\%on Yelp Full ordinal regression with 3\.2M parameters versus a 10\-model average of66\.35%66\.35\\%, and\+2\.32%\+2\.32\\%mean corruption accuracy on CIFAR\-10\-C\. All results with reportedpp\-values fall below theα=0\.05\\alpha=0\.05threshold after Bonferroni correction\. CR maintains a gradient tail ratioτ\\tau\(p99/mean\) of1\.011\.01–1\.021\.02against1\.071\.07–1\.091\.09for all typical activation functions baselines across every data fraction, a structural invariant we propose as the mechanistic driver of sample efficiency and a deployment\-time proxy for model reliability\.

Keywords:Sample Efficiency, Cross\-Domain Learning, Polynomial Neural Networks, Gradient Regularization, Reliable AI, Efficient Inference, Sentiment Analysis, Low\-Resource Learning

## 1Introduction

### 1\.1The Deployment Gap in Cross\-Domain Deep Learning

The prevailing research paradigm assumes that data is abundant, compute is elastic, and correctness can be evaluated against a fixed benchmark\. Production deployment assumes none of these things\. Across enterprise applications spanning tabular prediction, natural language processing, and computer vision, three structural constraints consistently limit what academic models can deliver in practice\.

Data scarcityLabeled data for specialized domains arrives slowly and expensively\. In medical classification, curated tabular datasets like Pima Diabetes reflect the realistic scale of institutional records available to most practitioners: hundreds of examples, not hundreds of thousands\[[2](https://arxiv.org/html/2605.24340#bib.bib2)\]\. In NLP, annotation of domain\-specific text for fine\-grained sentiment, compliance risk, or document routing costs $0\.50–$5\.00 per label with quality control\[[1](https://arxiv.org/html/2605.24340#bib.bib1)\], and model retraining is required on every distribution shift\. In vision, new product categories, manufacturing defects, and rare environmental conditions all suffer cold\-start data problems that standard benchmarks do not measure\. BERT\-scale models require tens of thousands of labeled examples to reliably fine\-tune for specialized domains\[[3](https://arxiv.org/html/2605.24340#bib.bib3)\]; annotation at that scale costs $10,000–$50,000 per vertical per update cycle\.

Reliability and model behaviorEnterprises deploying predictive models face a trust problem that benchmark accuracy does not resolve\. A model that achieves85%85\\%on a held\-out test set but produces extreme, unpredictable outputs on novel inputs is operationally unsafe\. The demand for interpretable and reliable AI in regulated industries is well documented\[[6](https://arxiv.org/html/2605.24340#bib.bib6),[7](https://arxiv.org/html/2605.24340#bib.bib7)\], yet the most widely deployed regularization strategies provide no mechanistic handle on gradient behavior at inference time\. Without such a handle, model audits are expensive and post\-hoc explanations are post\-hoc approximations\.

Compute and inference costThe compute cost of training large NLP models has grown by orders of magnitude faster than hardware efficiency gains\[[5](https://arxiv.org/html/2605.24340#bib.bib5)\], making continuous retraining in production financially unsustainable for most organizations\. Real\-time inference pipelines on commodity hardware cannot accommodate transformer\-scale models for every classification task\. Architectures that deliver competitive accuracy at 3–4M parameters, with inference\-cost derivative tracking and no iterative solver, have direct operational value across industries where latency and cost\-per\-inference are binding constraints\[[5](https://arxiv.org/html/2605.24340#bib.bib5)\]\.

CR is built around all three constraints simultaneously: a single inductive bias, DREG, that tightens gradient behavior, reduces the data required to reach a given accuracy level, and produces a measurable structural invariant that can be monitored in production\.

### 1\.2The Technical Gap

Existing regularization approaches do not address all three constraints jointly\. Dropout\[[10](https://arxiv.org/html/2605.24340#bib.bib10)\]disrupts co\-adaptation but leaves layer\-wise Jacobians unconstrained and provides no inference\-time reliability signal\. Weight decay\[[11](https://arxiv.org/html/2605.24340#bib.bib11)\]penalizes parameter magnitude, not sensitivity to input variation\. Spectral Normalization\[[12](https://arxiv.org/html/2605.24340#bib.bib12)\]provides global Lipschitz bounds but truncates expressive capacity uniformly, incurring an accuracy penalty\. Input Gradient Penalty\[[13](https://arxiv.org/html/2605.24340#bib.bib13)\]penalizes only the terminal Jacobian and requires double backpropagation at𝒪​\(3​Cf\)\\mathcal\{O\}\(3C\_\{f\}\)cost\. Sobolev Training\[[14](https://arxiv.org/html/2605.24340#bib.bib14)\]requires ground\-truth derivative labels unavailable in most practical settings\. Batch Normalization\[[15](https://arxiv.org/html/2605.24340#bib.bib15)\]stabilizes activations but does not constrain sensitivity layer\-wise\. The missing primitive is layer\-wise derivative control: a sensitivity budget enforced at each hidden layer, analytically, during the standard forward pass, without accuracy sacrifice or inference overhead\.

### 1\.3Contributions

1. 1\.ArchitectureChainzRulereplaces typical activation functions activations with learnable cubic polynomial layers\. A dual\-stream forward pass propagates both activations and their Jacobians simultaneously\. DREG penalizes the Frobenius norm of each layer’s Jacobian, imposing a per\-layer sensitivity budget at𝒪​\(d⋅Cf\)\\mathcal\{O\}\(d\\cdot C\_\{f\}\)cost with no second\-order graph\.
2. 2\.Cross\-domain accuracyCR achieves competitive or state\-of\-the\-art results across tabular \(Pima Diabetes\), NLP \(SST\-5, Yelp Full\), and vision \(CIFAR\-10\-C\) tasks, demonstrating that the polynomial\-DREG inductive bias generalizes across input modalities and task types without architecture\-level modification\.
3. 3\.Sample efficiencyCR leads all typical activation functions baselines at every tested data fraction \(5%–100%\) on Pima Diabetes and achieves statistically significant improvements on SST\-5 using approximately 5% of the training data used by the prior benchmark\.
4. 4\.Reliability and interpretabilityCR maintains gradient tail ratioτ∈\[1\.01,1\.02\]\\tau\\in\[1\.01,1\.02\]versus\[1\.07,1\.09\]\[1\.07,1\.09\]for all typical activation functions baselines at every data fraction \(p<0\.001p<0\.001\)\. This structural invariant is directly monitorable at inference time as a deployment reliability signal\.
5. 5\.Deployment evidenceSentivity AI runs CR in production for social media and market sentiment analysis, providing an existence proof that these properties transfer from benchmark to operational settings\.

## 2Architecture

### 2\.1PolyLayer

Each layer ofChainzRulereplaces the standard fixed activation function with a learnable cubic polynomial applied per neuron\. Rather than passing a pre\-activation through a hardcoded nonlinearity like typical activation functions, each neuron learns its own polynomial coefficients during training, giving the network the flexibility to approximate smooth, high\-order functions with far fewer layers and parameters than piecewise\-linear architectures require\. The cubic degree is fixed by grid search in prior work\[[17](https://arxiv.org/html/2605.24340#bib.bib17)\]\.

The practical upside over typical activation functions is smoothness\. Typical activation functions activations have a hard kink at zero: their derivative is undefined at that point and piecewise\-constant everywhere else\. This creates structural cliffs in the gradient landscape that generate spike events, particularly under small batches\. Polynomial activations are infinitely differentiable everywhere, which means the Jacobian can be tracked analytically through every layer without numerical instability, and gradient behavior becomes a first\-class observable rather than a diagnostic afterthought\.

### 2\.2Dual\-Stream Forward Pass and DREG

CR runs two streams simultaneously during the forward pass: a standard value stream producing predictions, and a derivative stream tracking how sensitive each layer’s output is to the original input\. The derivative stream is updated analytically at each layer using the chain rule, requiring no second backward pass and no second\-order computational graph\. The added cost is proportional to input dimensionality times the cost of a single forward pass, substantially cheaper than competing derivative\-aware methods such as Sobolev Training or Input Gradient Penalty, which require double backpropagation\[[14](https://arxiv.org/html/2605.24340#bib.bib14),[13](https://arxiv.org/html/2605.24340#bib.bib13)\]\.

Differential Regularization \(DREG\) uses that derivative stream to impose a per\-layer sensitivity budget during training\. At each layer, the magnitude of the tracked Jacobian is penalized directly in the loss\. This acts as a soft governor on gradient extremes: the network is discouraged from developing the heavy\-tailed sensitivity patterns that destabilize learning under scarce data, while retaining the expressive capacity to fit complex decision boundaries\. Unlike Spectral Normalization\[[12](https://arxiv.org/html/2605.24340#bib.bib12)\], which enforces a hard global Lipschitz bound and uniformly truncates the model’s representational capacity, DREG is selective: it suppresses tail events without flattening the features needed for discriminative accuracy\. The resulting model is not merely regularized but structurally auditable:τ\\taucan be computed on any batch at inference time, providing a live reliability signal without additional instrumentation\.

## 3Results

All experiments compare CR against Vanilla MLP, MLP \+ Dropout, and MLP \+ Weight Decay retrained under identical conditions, plus published baselines where available\. Statistical tests use paired one\-sidedtt\-tests and Wilcoxon signed\-rank tests with Bonferroni correction\. All results for whichpp\-values are reported fall below theα=0\.05\\alpha=0\.05significance threshold\.

#### Pima Diabetes \(Tabular\)

CR achieves85\.71%±2\.01%85\.71\\%\\pm 2\.01\\%\(6 seeds\), statistically superior to both XGBoost/RF \(p=0\.0047p=0\.0047\) and SVM \(p=0\.0039p=0\.0039\), and leads all typical activation functions baselines at every data fraction fromn=32n=32\(5%\) ton=614n=614\(100%\)\. The largest advantage occurs at 10% data \(\+3\.25%\+3\.25\\%over the best baseline\), the regime most analogous to cold\-start prediction in low\-resource production settings\. Prior best: XGBoost / RF at82\.35%82\.35\\%\[[18](https://arxiv.org/html/2605.24340#bib.bib18)\]; SVM at82\.20%82\.20\\%\[[19](https://arxiv.org/html/2605.24340#bib.bib19)\]\.

#### SST\-5 \(NLP, Frozen Encoder\)

On the frozen encoder task, CR \(46\.20%±0\.37%46\.20\\%\\pm 0\.37\\%, 4 seeds\) is statistically superior to the RNTN \(p=0\.0362p=0\.0362\) while using approximately 5% of RNTN’s phrase\-level training data, a data\-efficiency ratio of roughly20×20\\times\. Prior best \(frozen, same regime\): RNTN at45\.7%45\.7\\%\[[20](https://arxiv.org/html/2605.24340#bib.bib20)\]\.

#### SST\-5 \(NLP, Fine\-Tuned BERT\)

With a fine\-tuned BERT backbone\[[3](https://arxiv.org/html/2605.24340#bib.bib3)\], CR reaches55\.79%55\.79\\%versus Munikar et al\.’s BERT\-base linear head trained on18×18\\timesmore labeled sentences\[[21](https://arxiv.org/html/2605.24340#bib.bib21)\]\. This is a single run, directionally positive but not yet statistically confirmed; multi\-seed evaluation is in progress\. The result demonstrates that CR’s polynomial head integrates cleanly with large pretrained encoders and retains its advantage even when the backbone is unfrozen\.

#### Yelp Full \(NLP, Ordinal Regression\)

CR achieves70\.17%70\.17\\%with 3\.22M parameters on 750,000\-example ordinal sentiment regression, exceeding the average of ten published fixed\-embedding systems by3\.823\.82percentage points\. This is not a claim against transformer\-scale models, which reach84\.49%84\.49\\%\[[25](https://arxiv.org/html/2605.24340#bib.bib25)\], but within the fixed\-embedding tier CR leads all published baselines\. Notably, typical activation functions with DREG degrades to58\.98%58\.98\\%on this task while CR maintains70\.17%70\.17\\%, confirming that the polynomial substrate’sC∞C^\{\\infty\}smoothness is the operative ingredient on high\-dimensional continuous sentiment manifolds, not the regularization alone\. Prior best \(fixed embeddings\): VDCNN64\.72%64\.72\\%\[[23](https://arxiv.org/html/2605.24340#bib.bib23)\], fastText63\.90%63\.90\\%\[[22](https://arxiv.org/html/2605.24340#bib.bib22)\], Char\-CNN62\.33%62\.33\\%\[[24](https://arxiv.org/html/2605.24340#bib.bib24)\]; 10\-model average66\.35%66\.35\\%\.

#### CIFAR\-10\-C \(Vision\)

CR achieves\+2\.32%\+2\.32\\%mean corruption accuracy over the matched Vanilla MLP head \(t=4\.33t=4\.33,p=6\.18×10−5p=6\.18\\times 10^\{\-5\}, significant after Bonferroni correction\)\. Gains are largest on high\-frequency noise corruptions: Impulse\+5\.80%\+5\.80\\%, Gaussian\+4\.96%\+4\.96\\%, Glass Blur\+5\.23%\+5\.23\\%\. No corruption\-aware training is used; robustness is a structural byproduct of derivative control on clean data, not a result of augmentation\.

## 4Mechanistic Analysis: Gradient Tail Ratio

We operationalize gradient stability via thetail ratioτ=p99/mean\\tau=\\text\{p99\}/\\text\{mean\}of per\-sample input gradient norms\. Highτ\\tauindicates rare, extreme gradient events that dominate the update direction, exactly the failure mode that amplifies under data scarcity, when batches are small and noisy, and the mode most responsible for unpredictable model behavior at inference time on out\-of\-distribution inputs\[[16](https://arxiv.org/html/2605.24340#bib.bib16)\]\.

Across all Pima fractions \(6 seeds per fraction\), CR maintainsτ∈\[1\.01,1\.02\]\\tau\\in\[1\.01,1\.02\]while all typical activation function baselines maintainτ∈\[1\.07,1\.09\]\\tau\\in\[1\.07,1\.09\], an average gap of5\.995\.99percentage points \(p<0\.001p<0\.001against each baseline\)\. This finding extends and sharpens the Phase 1 result\[[17](https://arxiv.org/html/2605.24340#bib.bib17)\], where DREG reduced peak gradient volatility by23\.1%23\.1\\%across all model capacities on MNIST \(p<10−6p<10^\{\-6\}, 24/24 paired comparisons\), with the raw tail ratio of POLY DREG remaining stable at15\.2215\.22–19\.4319\.43while unregularized ReLU baselines escalated from17\.7917\.79to35\.6035\.60as capacity scaled\.

Critically,τ\\taufor CR is invariant to data fraction: the same suppression present atn=614n=614is present atn=32n=32\. This structural invariance distinguishes DREG from regularizers that improve stability only when data is abundant enough for stable convergence, and confirms that derivative control is a property of the architecture rather than an artifact of optimization under full data\.

The mechanism is twofold\. In typical activation functions networks, Jacobians are piecewise constant with cliffs atz=0z=0\. Under small batches from scarce data, activations frequently land in high\-sensitivity regions, generating tail events that distort gradient direction\. DREG penalizes‖S\(l\)‖F2\\\|S^\{\(l\)\}\\\|\_\{F\}^\{2\}layer\-by\-layer, suppressing these events before they compound across depth\. The polynomial substrate removes the structural source:ϕ′​\(z\)\\phi^\{\\prime\}\(z\)is analytic everywhere, so there are noz=0z=0cliffs from which Jacobian spikes can originate\.

The tail ratio has a direct deployment interpretation\. A production model withτ≈1\.01\\tau\\approx 1\.01produces output distributions that are structurally less sensitive to input perturbations than a model atτ≈1\.09\\tau\\approx 1\.09\. This translates to reduced false\-positive rates on novel inputs, new named entities, unseen product categories, and distribution\-shifted data, without requiring retraining or manual post\-filtering\. In this sense,τ\\tauis not merely a diagnostic but a deployable reliability metric\.

## 5Deployment and Business Case

Sentivity AIis a production market intelligence platform providing sentiment analysis on social media, news, and other digital mediums\. CR and DREG constitute the underlying classification engine in the production inference stack\. The Sentivity deployment is presented not as the primary motivation for CR but as an existence proof: a production system built on CR’s properties that validates their real\-world transferability\.

### 5\.1Why These Properties Matter in Deployment

The three axes CR optimizes, sample efficiency, structural reliability, and inference efficiency, map directly to the constraints that determine whether a model can be deployed and maintained in practice\.

Reduced annotation costProfessional\-grade domain annotation costs $1–$5 per label with inter\-annotator agreement checks\[[1](https://arxiv.org/html/2605.24340#bib.bib1)\]\. At 10,000 examples, that is $10,000–$50,000 per vertical per training run\. An architecture that achieves equivalent accuracy at 25% of the data reduces annotation cost by $7,500–$37,500 per run\. CR’s20×20\\timesdata\-efficiency ratio on SST\-5 and consistent leads at 5–25% Pima fractions translate directly to this reduction\.

Faster time\-to\-deployment on new tasksNew product categories, regulatory domains, and emerging topics all share a cold\-start problem: the model must generalize from a small labeled seed before a full annotation cycle can complete\. The Pima 5% result \(n=32n=32,\+1\.52%\+1\.52\\%over best baseline\) and SST\-5 frozen result \(20×20\\timesdata efficiency\) are the closest experimental analogues to this regime\. A model that reaches acceptable accuracy atn=50n=50rather thann=1000n=1000cuts cold\-start deployment time by weeks\.

Auditable behavior without post\-hoc approximationExplainability tools such as LIME\[[8](https://arxiv.org/html/2605.24340#bib.bib8)\]and SHAP\[[9](https://arxiv.org/html/2605.24340#bib.bib9)\]are post\-hoc approximations that may not faithfully represent model internals\[[7](https://arxiv.org/html/2605.24340#bib.bib7)\]\. CR’sτ\\tauis computed directly from the model’s own forward pass and reflects actual gradient behavior, not a surrogate\. For regulated industries and enterprise risk management, a first\-class reliability signal computed at inference cost is operationally more valuable than an approximate explanation computed separately\.

CPU\-viable inferenceCR’s 3\.2M parameter footprint and inference\-cost derivative tracking make it deployable on CPU\-only infrastructure\. Real\-time ingestion of high\-frequency data streams on commodity hardware cannot accommodate transformer inference at scale\. Green AI considerations\[[5](https://arxiv.org/html/2605.24340#bib.bib5)\]further favor lightweight architectures with competitive accuracy over large models that require GPU clusters for every inference call\.

### 5\.2Production Architecture

Sentivity AI’s inference pipeline applies CR as a sentiment scoring head on top of fixed sentence embeddings \(all\-mpnet\-base\-v2\), processing social media and financial text in real time\. The fixed\-embedding regime mirrors the SST\-5 frozen encoder experiments directly\. The ordinal Yelp Full results \(70\.17%70\.17\\%, fixed GloVe\) provide large\-scale validation that the architecture scales to 750,000\-example corpora under realistic NLP noise\. The gradient tail ratio is monitored in production as a reliability signal; no anomaly post\-filtering layer is required\.

## 6Limitations

SST\-5 BERT:The fine\-tuned BERT result \(55\.79%55\.79\\%\) is a single run and lacks statistical power\. Treat as directional\. Multi\-seed evaluation is in progress\.

Yelp protocol:CR uses fixed GloVe embeddings and ordinal regression; most Yelp Full baselines use categorical classification with fine\-tuned embeddings\. The70\.17%70\.17\\%result is not directly comparable to the 10\-model average in the same sense as the Pima comparisons\.

Pima scale:768 total samples is a small dataset\. Results are robust across 6 seeds and stratified splits, but broader tabular validation across additional medical and enterprise datasets would strengthen the generalization claim\.

Backbone integration:All experiments use CR as a classification head on a frozen or fine\-tuned backbone\. Whether DREG integrated directly into transformer attention layers provides consistent advantages is an open question for future work\.

Vision scope:CIFAR\-10\-C results use a matched MLP head, not a full convolutional architecture\. The\+2\.32%\+2\.32\\%corruption robustness result reflects CR’s behavior as a head component under distribution shift, not as a standalone vision backbone\.

## 7Conclusion

ChainzRuledemonstrates that layer\-wise derivative control is a tractable and effective inductive bias across tabular, NLP, and vision tasks\. A single architecture, a polynomial layer governed by DREG, simultaneously improves sample efficiency, structural reliability, and inference cost relative to typical activation functions baselines, without modifying the training procedure or adding inference\-time overhead beyond a single forward pass\.

The gradient tail ratioτ\\tauis a measurable, structurally stable proxy for this control: CR maintainsτ≈1\.01\\tau\\approx 1\.01–1\.021\.02against1\.071\.07–1\.091\.09for all typical activation functions baselines at every data fraction tested, with statistical significance atp<0\.001p<0\.001throughout\. This invariant is not merely a training diagnostic; it is a live deployment signal that can be monitored without additional instrumentation\.

For any domain where annotation cost, model reliability, and inference budget are binding constraints, CR offers a deployable path to competitive accuracy at significantly reduced data and compute requirements\. Annotation cost reductions of4×4\\times–20×20\\timesare empirically supported\. Gradient stability under distribution shift reduces false\-positive rates in streaming inference\. A 3\.2M parameter footprint makes CPU deployment viable\. Sentivity AI’s production deployment confirms these properties transfer from benchmark to operational settings\. A more complete theoretical treatment, including formal convergence analysis, bounds on sample complexity under DREG, and transformer integration, is the subject of ongoing work\.

## References

## References

- \[1\]A\. Ratner, C\. De Sa, S\. Wu, D\. Selsam, and C\. Ré, “Data Programming: Creating Large Training Sets, Quickly,”Advances in Neural Information Processing Systems \(NeurIPS\), 2016\.
- \[2\]D\. Dua and C\. Graff, “UCI Machine Learning Repository,” University of California, Irvine, 2017\.[https://archive\.ics\.uci\.edu/ml](https://archive.ics.uci.edu/ml)
- \[3\]J\. Devlin, M\.\-W\. Chang, K\. Lee, and K\. Toutanova, “BERT: Pre\-training of Deep Bidirectional Transformers for Language Understanding,”Proceedings of NAACL\-HLT, pp\. 4171–4186, 2019\.
- \[4\]E\. Strubell, A\. Ganesh, and A\. McCallum, “Energy and Policy Considerations for Deep Learning in NLP,”Proceedings of ACL, 2019\.
- \[5\]R\. Schwartz, J\. Dodge, N\. A\. Smith, and O\. Etzioni, “Green AI,”Communications of the ACM, vol\. 63, no\. 12, pp\. 54–63, 2020\.
- \[6\]F\. Doshi\-Velez and B\. Kim, “Towards a Rigorous Science of Interpretable Machine Learning,”arXiv:1702\.08608, 2017\.
- \[7\]C\. Rudin, “Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead,”Nature Machine Intelligence, vol\. 1, pp\. 206–215, 2019\.
- \[8\]M\. T\. Ribeiro, S\. Singh, and C\. Guestrin, “Why Should I Trust You?: Explaining the Predictions of Any Classifier,”Proceedings of KDD, pp\. 1135–1144, 2016\.
- \[9\]S\. M\. Lundberg and S\.\-I\. Lee, “A Unified Approach to Interpreting Model Predictions,”Advances in Neural Information Processing Systems \(NeurIPS\), 2017\.
- \[10\]N\. Srivastava, G\. Hinton, A\. Krizhevsky, I\. Sutskever, and R\. Salakhutdinov, “Dropout: A Simple Way to Prevent Neural Networks from Overfitting,”Journal of Machine Learning Research, vol\. 15, pp\. 1929–1958, 2014\.
- \[11\]I\. Loshchilov and F\. Hutter, “Decoupled Weight Decay Regularization,”ICLR, 2019\.
- \[12\]T\. Miyato, T\. Kataoka, M\. Koyama, and Y\. Yoshida, “Spectral Normalization for Generative Adversarial Networks,”arXiv:1802\.05957, 2018\.
- \[13\]H\. Drucker and Y\. Le Cun, “Improving Generalization Performance using Double Backpropagation,”IEEE Transactions on Neural Networks, vol\. 3, no\. 6, pp\. 991–997, 1992\.
- \[14\]W\. M\. Czarnecki, S\. Osindero, M\. Jaderberg, G\. Swirszcz, and R\. Pascanu, “Sobolev Training for Neural Networks,”arXiv:1706\.04859, 2017\.
- \[15\]S\. Ioffe and C\. Szegedy, “Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,”Proceedings of ICML, pp\. 448–456, 2015\.
- \[16\]S\. Fort, P\. Hu, and B\. Lakshminarayanan, “Deep Ensembles: A Loss Landscape Perspective,”arXiv:1912\.02757, 2019\.
- \[17\]R\. Martnishn and S\. Anderson, “Layer\-wise Derivative Controlled Networks,”arXiv preprint, Sentivity AI / Virginia Tech, 2025\.
- \[18\]O\. Yangin, “Gradient Boosting Methods for Disease Prediction,” Master’s Thesis, 2019\. Handle:hdl\.handle\.net/20\.500\.14124/1152
- \[19\]S\. Karatsiolis and C\. N\. Schizas, “Region Based Support Vector Machine Algorithm for Medical Diagnosis on Pima Indian Diabetes Dataset,”Proceedings of IEEE EMBC, 2012\.
- \[20\]R\. Socheret al, “Recursive Deep Models for Semantic Compositionality over a Sentiment Treebank,”Proceedings of EMNLP, pp\. 1631–1642, 2013\.
- \[21\]M\. Munikar, S\. Shakya, and A\. Shrestha, “Fine\-Grained Sentiment Classification using BERT,”arXiv:1910\.03474, 2019\.
- \[22\]A\. Joulin, E\. Grave, P\. Bojanowski, and T\. Mikolov, “Bag of Tricks for Efficient Text Classification,”Proceedings of EACL, 2017\.
- \[23\]A\. Conneau, H\. Schwenk, L\. Barrault, and Y\. Lecun, “Very Deep Convolutional Networks for Text Classification,”Proceedings of EACL, 2017\.
- \[24\]X\. Zhang, J\. Zhao, and Y\. LeCun, “Character\-level Convolutional Networks for Text Classification,”NeurIPS, 2015\.
- \[25\]M\. Zulqarnainet al, “An Improved Gated Recurrent Unit Based on Auto Encoder for Sentiment Analysis,”International Journal of Information Technology, vol\. 15, no\. 1, pp\. 587–599, 2023\.
- \[26\]D\. Hendrycks and T\. Dietterich, “Benchmarking Neural Network Robustness to Common Corruptions and Perturbations,”ICLR, 2019\.