Chapter Four · failure evidence
What Regularization Methods got wrong, from 84 dissertations
Across numerous applications, regularization techniques frequently impaired predictive accuracy, disrupted optimization, or introduced severe parameter estimation bias. Practitioners repeatedly found that methods such as Lasso, Ridge, dropout, and spatial penalties underperformed simpler unregularized baselines or induced pathological failures like mode collapse and over-smoothing. These records come from PhD theses at 25 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
L1 and Lasso penalties cause estimation bias, arbitrary feature elimination, and instability under collinearity
Applying L1 and Lasso penalties frequently introduced severe shrinkage bias and degraded predictive accuracy by penalizing or completely eliminating important predictor coefficients. In several settings, L1 regularization artificially split effects across correlated markers, failed to respect spatial or global feature structures, or required separate fine-tuning phases that made alternative formulations preferable.
Tried and failed
Lasso regularization applied to multi-state survival models. Outcome: worse than baseline. Reason: produced overly sparse models due to elevated regularization-induced bias
Methods for Flexible Survival Analysis and Prediction of Semi-Competing Risks · Harvard
Tried and failed
L1 and elastic net regularization on LSTM applied to time series electrical load forecasting. Outcome: worse than baseline. Reason: L1 penalty degraded prediction accuracy compared to unregularized or pure L2 regularization
A Deep Learning-based Dynamic Demand Response Framework · Virginia Tech
Tried and failed
L1 regularization for sparse regression applied to symbolic differential equation identification. Outcome: worse than baseline. Reason: regularization strength needed for sparsity degraded model prediction accuracy
Interpretable Physics-informed Machine Learning Methods for Scientific Modeling and Data Analysis · MIT
Tried and failed
Lasso L1 regularization applied to QTL mapping with linked genetic markers. Outcome: worse than baseline. Reason: High collinearity between linked loci caused effect sizes to be artificially split across markers
The structure of fitness landscapes across genotypes and environments · Harvard
Tried and failed
L1 regularization for write-aware neural network training applied to hardware-aware neural network optimization. Outcome: worse than baseline. Reason: Severely degraded model expressivity and caused hypersensitivity to regularization strength compared to L2
Co-architecting scalable intelligence: from hardware substrates to algorithms · UT Austin
Tried and failed
Node-level L1 and group Lasso regularization applied to tree ensembles for feature selection. Outcome: worse than baseline. Reason: L1 and group Lasso fail to enforce coordinated global feature selection across ensemble trees compared to L0-L2 penalties
Nonparametric High-dimensional Models: Sparsity, Efficiency, Interpretability · MIT
Tried and failed
Lasso regression with molecular descriptors and fingerprints applied to predicting organometallic complex emission energy. Outcome: no signal. Reason: L1 regularization failed to identify a predictive linear relationship between descriptors and emission energy
Lost to a baseline
Lasso regularized regression was outperformed by bias-corrected Lasso, bias-corrected ridge, and SEMMS across varying noise levels and sample sizes in recovering true ODE terms (Van der Pol and 2D spiral systems) and PDE terms (viscous Burgers' and 2D heat equations) due to Lasso selecting spurious higher-order terms.
Parameter estimation and inference for nonlinear dynamical systems · Cornell
Considered and rejected
Considered and rejected: Rejected standard L1 (Lasso) regularization because it causes parameter shrinkage and requires a separate unregularized fine-tuning phase.
Informed Equation Learning · Publikationssystem UB Tuebingen
Considered and rejected
Considered and rejected: Rejected standard Lasso L1 regularization in favor of Fused Lasso, because the fused lasso properly accounts for spatial ordering and natural continuity of genomic coordinates in tiling screen data.
Computational and experimental methods for CRISPR-based saturation mutagenesis screens · MIT
Considered and rejected
Considered and rejected: Rejected standard LASSO/BPDN magnitude regularization for parameter estimation because L1 penalties inherently bias and shrink true nonzero coefficients.
Physics-Inspired Machine Learning of Partial Differential Equations · Georgia Tech
Considered and rejected
Considered and rejected: Rejected Lasso regression in favor of Ridge regression because L1 regularization eliminates regressors (sparseness), creating problems when comparing variance decomposition across variable sets.
INFERENCE OF REPRESENTATIONS THROUGH STRUCTURE: REVISITING MARR’S TRI-LEVEL HYPOTHESIS OF NEUROSCIENCE · ScholarlyCommons at Penn
Considered and rejected
Considered and rejected: Decided against LASSO regression because L1 regularization shrinks weights to zero and excludes genes, preventing evaluation of full-transcriptome weight shifts across conditions.
Considered and rejected
Considered and rejected: Rejected using L1 Lasso regularization in the CDR clustering formulation because Ridge regression provides a closed-form solution enabling an efficient and scalable Convex Integer Program reformulation.
Advanced Data Analytics for Quality Assurance of Smart Additive Manufacturing · Virginia Tech
Considered and rejected
Considered and rejected: Rejected l1-regularized sparse regression methods (e.g., LASSO), probabilistic approaches, and deep learning baselines for system identification benchmarking due to unsuitability for exact variable selection.
Towards an Artificial Neuroscience: Analytics for Language Model Interpretability · MIT
Considered and rejected
Considered and rejected: Rejected L1 regularization in final retraining pipelines because L2 regularization produced superior accuracy after retraining
Hardware-Friendly Model Compression techniques for Deep Learning Accelerators · Georgia Tech
Considered and rejected
Considered and rejected: Rejected Lasso (L1) regularization in favor of Ridge (L2) regression to prevent penalizing variable coefficients to zero after multistage backward selection.
Predictive Models for Pediatric Cardiac Surgery Outcomes · Harvard
Dropout and stochastic regularizers degrade network capacity and disrupt training dynamics
Dropout and related stochastic masking techniques often degraded performance by excessively reducing effective model capacity on small architectures, physical modeling tasks, and large datasets where regularization was unneeded. These methods also disrupted numerical derivative approximations in differential equation learning, slowed single-epoch pre-training, and exacerbated gradient conflicts during supernet training.
Tried and failed
temporal activation-dependent dynamic dropout regularization applied to recurrent neural network training. Outcome: worse than baseline. Reason: failed to induce neuronal specialization and underperformed standard dropout
Tried and failed
adding batch normalization and dropout after layers applied to low-complexity neural network architectures. Outcome: worse than baseline. Reason: excessive regularization degraded performance on simpler models
Leveraging data characteristics for bug localization in deep learning programs · Iowa State
Tried and failed
dropout regularization in residual networks applied to image classification on small datasets. Outcome: worse than baseline. Reason: severely degraded training performance of the base classifier
Seen/Unseen classification using feature vector analysis · Iowa State
Tried and failed
dropout regularization applied to sliding window CNN regression. Outcome: worse than baseline. Reason: None
Tried and failed
regularization and hyperparameter tuning of neural networks applied to geometric deflection prediction. Outcome: worse than baseline. Reason: L1 regularization and dropout degraded prediction accuracy relative to baseline model
Tried and failed
dropout regularization in neural networks applied to learning differential equations from data. Outcome: worse than baseline. Reason: disrupted derivative approximation and prevented convergence to the underlying governing equation
Learning Differential Equations from Noisy, Limited Data · Cornell
Tried and failed
dropout and batch normalization applied to shallow MLP regression with small batches. Outcome: worse than baseline. Reason: regularization techniques degraded performance in low-capacity shallow networks trained on small batches
Tried and failed
dropout and L2 weight regularization applied to fluid dynamic drag force prediction. Outcome: worse than baseline. Reason: standard regularizers reduced model capacity needed to capture complex non-linear physical interactions
Science Guided Machine Learning: Incorporating Scientific Domain Knowledge for Learning Under Data Paucity and Noisy Contexts · Virginia Tech
Tried and failed
dropout and regularization in single-epoch training applied to foundation model pre-training. Outcome: worse than baseline. Reason: slowed training speed and degraded loss curves under non-repeating data regimes
Improving Foundation Models · Georgia Tech
Tried and failed
applying dropout regularization to recurrent neural networks applied to time series prediction on large datasets. Outcome: worse than baseline. Reason: regularization degraded performance without improving generalization due to the sufficiently large training dataset size
Traffic Signal Phase and Timing Prediction: A Machine Learning and Controller Logic Hybrid Approach · Virginia Tech
Lost to a baseline
In the 5-layer CNN architecture, applying Dropout yielded lower accuracy (54.25% at 50 epochs, 61.72% at 100 epochs) than the unregularized 4-layer baseline (61.25% at 50 epochs).
Advanced techniques for characterizing and predicting asphalt macrotexture Integrated Digital Techniques for Asphalt Macrotexture Characterization: From X-ray CT to Virtual Modeling and Deep Learning for Efficient, Low-Cost Assessment · University of Nottingham Repository
Considered and rejected
Considered and rejected: Rejected using Drop-path regularization on small ShuffleNASNet-A models (<1M parameters) as it degraded performance unless stabilized by BatchNorm.
Improving the automated search of neural network architectures · Publikationssystem UB Tuebingen
Considered and rejected
Considered and rejected: Decided against strong data augmentations/regularizations (DropConnect, dropout, weight decay) in ViT supernet training because they exacerbate gradient conflicts.
Machine learning meets and enhances protein engineering · UT Austin
L2 regularization and weight decay cause underfitting and fail on structured or sparse data
Adding L2 penalties or weight decay often degraded accuracy compared to unpenalized models and led to severe underfitting in autoencoders and neural networks. Ridge regularization also underperformed when dealing with sparse web datasets or severe multicollinearity where sparse selection methods like Lasso or group Lasso were required.
Tried and failed
weight decay regularization in loss function applied to fabric deformation prediction model. Outcome: worse than baseline. Reason: non-zero weight decay penalty reduced model performance compared to zero regularization
Wrinkling behaviour of biaxial non-crimp fabrics during preforming · Cambridge
Considered and rejected
Considered and rejected: Rejected L2/dropout regularization on small subsets for learning raw audio features because it degraded classification performance compared to scaling data volume.
Leveraging Generative Models for Music and Signal Processing · ResearchWorks
Tried and failed
L1 and L2 regularized non-negative linear regression applied to bulk gene expression deconvolution. Reason: Regularization penalties showed no performance improvement over standard non-negative regression.
Tried and failed
GLM with L2 ridge regularization applied to neural calcium activity encoding models. Outcome: worse than baseline. Reason: L2 regularization underperformed compared to group lasso in predicting test deviance
Organization of Neural Representations in Mouse Posterior Cortex for Dynamic Navigation Decisions · Harvard
Tried and failed
L2 regularization on autoencoder dimensionality reduction applied to unsupervised anomaly detection. Reason: higher regularization penalties caused underfitting and unpredictable performance degradation
On the Effectiveness of Dimensionality Reduction for Unsupervised Structural Health Monitoring Anomaly Detection · Virginia Tech
Tried and failed
L2 and elastic net regularization applied to recurrent neural network hyperparameter tuning. Outcome: worse than baseline. Reason: Failed to yield satisfactory predictive performance compared to small L1 regularization.
Blockchain-based Peer-to-peer Electricity Trading Framework Through Machine Learning-based Anomaly Detection Technique · Virginia Tech
Tried and failed
L1 and L2 regularized linear regression applied to cluster expansion effective cluster interactions. Outcome: worse than baseline. Reason: regularization penalised important cluster interactions, degrading energy prediction accuracy compared to unregularized fits
Order Under Pressure: Structural and Magnetic Characterization at Extreme Stresses · MIT
Lost to a baseline
Ridge regression (L2 regularization) achieved lower prediction performance than LASSO regression across voxels when modeling fMRI responses.
COMPUTATIONAL MODELS OF FEATURE REPRESENTATIONS IN THE VENTRAL VISUAL STREAM · JScholarship
Considered and rejected
Considered and rejected: L2-regularized logistic regression (ridge) was rejected in favor of L1-regularization (LASSO) because L1 yielded identical accuracy while providing more interpretable, sparse feature selection
Soil Microbial Assembly and Function in Agroecosystems Under Various Management and Disturbance Regimes · Scholars' Bank
Considered and rejected
Considered and rejected: Rejected L2 regularization (Ridge regression) in favor of L1 regularization (Lasso) for logistic regression to prioritize model simplicity and generalizability
Considered and rejected
Considered and rejected: Rejected Ridge regression and standard multiple linear regression in favor of LASSO (L1 regularization) to handle severe multicollinearity among co-regulated transcripts.
Transcriptomics in pulmonary arterial hypertension - diagnostics and pathobiology · Imperial
Considered and rejected
Considered and rejected: Rejected relying solely on L2 ridge regularization for sparse web datasets because it causes performance degradation.
Tensor Decomposition Method Applied to Recommendation Systems · Queens University Institutional Repository
Lost to a baseline
Neural network models (SNN and DNN) with L2 regularization underperformed simpler SVR and RFR models in generalizability on the unseen test dataset (DNN test MSE was nearly double that of RFR).
Towards Born Qualification of AM Components: High Temperature Fatigue Testing and Microstructural Characterization of AM IN718 · Georgia Tech
Spatial, temporal, and variational regularizers cause distortion, oversegmentation, and smoothing artifacts
Continuous and spatial regularization penalties frequently introduced artificial visual priors, rounded off sharp geometric edges, or produced staircase artifacts across image and shape reconstructions. In time-series and inverse problems, weak or excessive smoothing caused oversegmentation, introduced systematic bias across temporal sequences, or failed to outperform unregularized baselines.
Tried and failed
phase derivative regularization loss applied to speech enhancement without clean magnitude. Outcome: worse than baseline. Reason: did not outperform direct phase loss when clean magnitude is unknown
Incorporating Geometric and Consistency Constraints into Deep Models for Robust Phase Reconstruction and Speech Enhancement · Georgia Tech
Tried and failed
augmentation consistency regularization with hard thresholding applied to semi-supervised medical image segmentation. Reason: hard sample rejection limits the regularization benefit compared to soft adaptive sample reweighting
Randomized dimension reduction with statistical guarantees · UT Austin
Tried and failed
first-order variational regularization applied to image segmentation with smooth intensity gradients. Reason: causes staircase over-segmentation artifacts on steep gradients and cannot detect crease discontinuities
Variational methods and its applications to computer vision · Imperial
Tried and failed
standard regularization to mitigate deep model overfitting applied to image-based parameter regression. Outcome: overfit. Reason: standard regularizers caused underfitting or failed to resolve ill-posed single-channel input ambiguities
AI/DEEP LEARNING WAVEFRONT SENSING FOR HEL USING TARGET IMAGE TO SIMPLIFY ADAPTIVE OPTICS SYSTEMS · Calhoun
Tried and failed
change-point segmentation with weak regularization applied to multivariate time series forecasting. Outcome: worse than baseline. Reason: low regularization causes oversegmentation, reducing sample size per segment and degrading predictive accuracy
Tried and failed
4-connected spatial regularization applied to dual-energy radiography material decomposition. Outcome: worse than baseline. Reason: regularization did not improve noisy reconstruction and optimal tuning collapsed back to unregularized segment-wise estimation
Reconstructing the Atomic Number of Cargo X-ray Images using Dual Energy Radiography · MIT
Tried and failed
pixel-space regularization in activation maximization applied to visual feature attribution and visualization. Reason: produces faint high-frequency patterns across the image rather than localized semantic edits
Tried and failed
isotropic G-norm regularization on signed distance fields applied to geometric shape inpainting and denoising. Outcome: worse than baseline. Reason: lack of directional guidance rounded sharp features and produced ill-conditioned fourth-order PDEs
Geometric representation of multi-dimensional data and its applications · Georgia Tech
Lost to a baseline
In X-ray tomography with 1% noise and active subspace dimension r = 1, TSVD (84.58% relative error) outperformed Tikhonov (21.36% relative error), DI (21.36% relative error), and DIAS (21.35% relative error) under over-regularization (alpha = 10^8) where Tikhonov reached 80.77%, DI reached 78.08%, and DIAS reached 74.29%.
Accelerating inverse solutions with machine learning and randomization · UT Austin
Considered and rejected
Considered and rejected: Rejected using regularization during feature activation optimization because it introduces artificial visual priors that obscure true model mechanics
Machine Learning Beyond Accuracy: A Features Perspective On Model Generalization · MIT
Tried and failed
standard perceptual and spatial regularization losses applied to diffusion model latent optimization. Reason: regularizers did not improve generated image quality during latent space optimization
Advancing channel coding via deep learning · UT Austin
Tried and failed
regularization to smooth transient simulation artifacts applied to time-domain grey-box parameter estimation. Reason: regularization introduced systematic bias and degraded fit quality across the rest of the time series
Methods for Parameter Estimation with Devices in Microgrids · MIT
Tried and failed
temporal regularization using past trend estimates applied to spatiotemporal functional stream decomposition. Outcome: unstable. Reason: using estimated values instead of raw observations caused estimation errors to accumulate over time
Novel Learning Methods for High-dimensional Data with Applications in Process Modeling and Monitoring · Georgia Tech
Advanced, spectral, and adversarial regularizers trigger optimization instability and model collapse
Specialized regularizers such as direct Lipschitz constraints and entropy penalties triggered training collapse or mode collapse in generative adversarial networks. Other advanced methods suffered from spectral constraints losing to simple baselines, batch normalization interfering with gradient penalties, or graph penalties overly constraining representation capacity.
Lost to a baseline
Weighted BCE (T=1.5) regularized student lost to Plain BCE (T=1.0) and suffered training collapse (validation soft BCE 0.2741 vs. 0.1526).
Goal-Conditioned Evaluation of Sustainable Development Goal Contributions in Theses and Dissertations · Virginia Tech
Tried and failed
gradient-based learning of regularization weight via reconstruction loss applied to unrolled optimization algorithms for sparse coding. Outcome: did not converge. Reason: reconstruction loss drives the regularization strength to zero to minimize unpenalized fitting error
Deep Learning for Inverse Problems in Engineering and Science · Harvard
Tried and failed
Direct Lipschitz regularization during training applied to adversarially robust neural networks. Outcome: worse than baseline. Reason: Caused collapse to constant classifiers or failed to yield certifiable models compared to layerwise normalization
Character-level Adversarial Robustness in Natural Language Processing · EPFL
Tried and failed
entropy-regularized generative adversarial networks applied to stochastic process approximation. Outcome: worse than baseline. Reason: standard regularization parameter led to severe mode collapse and high reverse KL divergence
Deep Learning And Uncertainty Quantification: Methodologies And Applications · Penn
Lost to a baseline
On MNIST (2*1k feed-forward), spectral regularized network achieved 98.66% test accuracy, losing to the dropout baseline at 98.70%.
Robust Submodular Partitioning and Linear Models of Deep ReLU Networks · ResearchWorks
Lost to a baseline
On CIFAR-10, spectral regularized feed-forward (3*4k) achieved 57.43% test accuracy, losing to dropout baseline at 57.50%.
Robust Submodular Partitioning and Linear Models of Deep ReLU Networks · ResearchWorks
Considered and rejected
Considered and rejected: Rejected batch normalization in waveform/power critics because it interferes with the WGAN gradient penalty calculation.
Leveraging audio-visual speech effectively via deep learning · Imperial
Considered and rejected
Considered and rejected: Rejected batch normalization in the critic for WGAN-GP because it invalidates the gradient penalty calculation.
A deep learning approach to wireless system design for channel sensing, contention & estimation · UT Austin
Considered and rejected
Considered and rejected: Rejected setting V > 0.8 in GAGAN regularization because it yielded no significant improvement in discriminator performance.
Human-controllable and structured deep generative models · Imperial
Tried and failed
excessive graph-based regularization penalty applied to continual learning for image prediction. Outcome: worse than baseline. Reason: over-regularizing learned edge distributions constrained model capacity and degraded predictive performance
Deep Probabilistic Models for Sequential Prediction · Cornell
Tried and failed
increasing hyperedge cardinality and graph sparsity regularization applied to hypergraph neural network time-series forecasting. Outcome: worse than baseline. Reason: Excessively large hyperedges and over-regularization degraded predictive performance and model representation capacity.
Graph-based Time-series Forecasting in Deep Learning · Virginia Tech
Considered and rejected
Considered and rejected: Rejected naive deep neural networks for learned regularization because large parameter counts require far more computation than standard Krylov solvers and destabilize across recurrent unrolling steps.
Scalar Scattering Theory and Physics-inspired Optimization for Computational Imaging · MIT
Regularized models fail to improve upon simpler or unregularized baselines
Regularized regression models and support vector machines frequently achieved worse or negligible performance gains compared to plain unregularized linear regressions, logistic models, or generalized additive models. In other instances, regularized approaches failed to prevent test error inflation, underperformed simple extrapolation, or fell short of straightforward domain baselines.
Tried and failed
Lasso regression for high-dimensional feature selection applied to time-series accelerometer feature subsets. Outcome: overfit. Reason: Unconstrained regularization retained too many predictors, inflating test mean squared error
Lost to a baseline
Elastic Net, Lasso, and Ridge regression (RMSE 1.08 across mono, dual, and triple therapies) were outperformed by unregularized stepwise linear regression (RMSE 1.0734 to 1.0758)
Predicting the Effects of Sedative Infusion on Acute Traumatic Brain Injury Patients · Virginia Tech
Lost to a baseline
Unregularized logistic regression achieved slightly higher training accuracy (83.5 ± 0.2% vs 83.1 ± 0.1%), training recall (80.1 ± 0.2% vs 78.3 ± 0.3%), and test F1 score (82.5 ± 1.5% vs 81.6 ± 1.1%) compared to soft-margin linear SVM.
Learning Simple Chemical Heuristics to Model and Discover Materials · MIT
Lost to a baseline
Least-squares and sparse least-squares (L1-regularized) polynomial surrogate models underperformed HierGP despite utilizing identical perfectly specified bases.
Lost to a baseline
Regularized Z optimization yielded equal or worse error than the non-regularized reference segment method at realistic 10% noise levels.
Reconstructing the Atomic Number of Cargo X-ray Images using Dual Energy Radiography · MIT
Lost to a baseline
Deep learning triple therapy (RMSE 1.0725) and regularized regressions (RMSE 1.08) achieved negligible or worse performance compared to standard linear regression (RMSE 1.0734) and GAM (RMSE 1.0734)
Predicting the Effects of Sedative Infusion on Acute Traumatic Brain Injury Patients · Virginia Tech
Tried and failed
ODE-based synthetic data generation and regularization applied to epidemic trend prediction. Outcome: worse than baseline. Reason: failed to produce well-correlated trend predictions compared to gradient matching
Artificial Intelligence for Data-centric Surveillance and Forecasting of Epidemics · Georgia Tech
Tried and failed
theory-informed variable selection with regularized regression applied to clinical treatment response prediction. Outcome: worse than baseline. Reason: models based on prior univariate literature underperformed simple baseline symptom-severity models
APPLYING MATHEMATICAL MODELS TO IMPROVE CLINICAL EVALUATION AND PREDICTION · Penn
Lost to a baseline
Simple extrapolation produced negative relative performance (-70.92% change) on least squares with Huber regularization on random data at N=1000.
On Large-Scale Optimization: Optimal Methods and Computer-Assisted Algorithm Design · JScholarship
Left open by the authors
Problems the authors named and did not get to.
Left open
Evaluate regularization methods on CTC loss to mitigate peakiness and improve adversarial robustness in hybrid ASR systems. Blocker: None
Understanding, Fortifying and Democratizing AI Security · Georgia Tech
Left open
Implement LASSO and Ridge regularization natively within the GSMP-VGLM coordinate-wise gradient descent optimization framework to prevent overfitting. Blocker: None
Left open
Learn a transformed parameter space using L1 regularization on gradient vectors to enforce gradient sparsity in zeroth-order optimization. Blocker: None
A NEW ZEROTH-ORDER ORACLE FOR DISTRIBUTED AND NON-STATIONARY LEARNING · DukeSpace
Left open
Apply regularized regressions (LASSO/Ridge) and Kernel PCA with train/test validation to distribution utility financial performance data. Blocker: None
Essays on Environmental, Fiscal, and Demand-Side Policies in Electricity Markets · Georgia Tech
Left open
Extend CLUSSO and Random CLUSSO algorithms to unbalanced multi-dimensional tensor regression using sparse regularized low-rank penalties. Blocker: None
STATISTICAL METHODS FOR VARIABLE SELECTION AND PREDICTION WITH PATHOMIC FEATURES · Penn
Left open
Develop inference procedures and Lasso-type regularized estimation methods for high-dimensional tensor regression models under weak L2-mixingale dependence. Blocker: None
On Estimation Methods in Tensor Regression Models · Scholarship at UWindsor Institutional Repository
Left open
Implement and evaluate class-dependent adaptive Lasso regularizations with penalty parameters conditioned on class probabilities. Blocker: None
Contributions to Efficient Statistical Modeling of Complex Data with Temporal Structures · Virginia Tech
Left open
Formulate and implement robust estimators and regularization penalties like LASSO and Elastic Net within the mixed-integer conic GLM optimization framework. Blocker: None
An optimization approach to generalized linear models · DSpace-CRIS at TU Wien
Left open
Develop and test regularization strategies to prevent overfitting in mismatched linear regression when the search radius is set too large. Blocker: None
Large-Scale Algorithms for Machine Learning: Efficiency, Estimation Errors, and Beyond · MIT
Left open
Develop regularization strategies to prevent overfitting in grammatical evolution feature extraction for time series classification. Blocker: None
One-Class Time Series Classification · Research Repository UCD
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.