Skip to content

Regularization Across Models — Cross-Topic Synthesis

Every regularizer is the same idea wearing a different costume: control effective capacity so the model fits signal, not sample. See INDEX.md for the full curriculum index.


Overview

Regularization controls effective model capacity; it is broader than adding a penalty to an objective. Penalties, constraints, noise injection, averaging, and stopping early all restrict which functions the training procedure can actually reach — and that is the common definition this page uses. Two equivalent lenses organize everything below.

One Frame, Two Lenses

Lens 1 — capacity control. For suitable regularity conditions, a penalized problem

\[ \min_\theta L(\theta)+\lambda R(\theta) \]

corresponds to a constrained problem \(\min_\theta L(\theta)\) subject to \(R(\theta)\le t\): shrinking \(t\) shrinks the hypothesis space, trading variance for bias (see Bias–Variance; the region-shape geometry is in Geometry of ML). The mapping between \(\lambda\) and \(t\) need not be one-to-one in degenerate or inactive cases.

Lens 2 — MAP priors. If the loss is a negative log-likelihood, the penalty is a negative log-prior:

\[ \hat{\theta}_{\text{MAP}} =\arg\max_\theta\;\log p(\mathcal{D}\mid\theta)+\log p(\theta) =\arg\min_\theta\;L(\theta)-\log p(\theta). \]
Penalty Prior on \(\theta_j\) Correspondence
\(\lambda\Vert\theta\Vert_2^2\) Gaussian \(\mathcal{N}(0,\tau^2)\) \(\lambda=\sigma^2/(2\tau^2)\) for Gaussian noise variance \(\sigma^2\)
\(\lambda\Vert\theta\Vert_1\) Laplace\((0,b)\) \(\lambda=\sigma^2/b\); sharp peak at 0 → exact zeros

Tighter prior (smaller \(\tau\) or \(b\)) = stronger penalty = smaller reachable region: the two lenses are one statement. Full derivations: Regularization theory; the MLE/MAP framing is in Probabilistic View of ML.


Mechanism Map

Mechanism Primary effect Models Failure mode
\(\ell_2\) penalty Shrinks parameters smoothly Linear models, SVM, neural networks Can retain many weak irrelevant features
\(\ell_1\) penalty Encourages exact zeros Linear and generalized linear models Unstable selection among correlated features
Elastic Net Combines shrinkage and sparsity Linear models Adds another tuning dimension
Tree depth / leaf constraints Limits partition complexity Trees, forests, boosting Excessive limits create high bias
Cost-complexity pruning Removes weak subtrees post hoc Decision trees Pruning parameter needs validation
Bagging Averages away variance Forests, any high-variance learner Cannot reduce bias; correlated members limit gains
Early stopping Restricts optimization trajectory Boosting, neural networks Requires a leakage-free validation rule
Weight decay \(\ell_2\)-style shrinkage during updates Deep networks Not identical to \(\ell_2\) loss penalty under Adam
Dropout Stochastic activation masking Neural networks Not equivalent to simple \(\ell_2\) in general
Data augmentation Encodes invariances through examples CNNs and other neural models Invalid transformations inject label noise
Label smoothing Softens targets, caps confidence Classifiers with cross-entropy Distorts calibration and distillation targets

Per-Family Instances

Linear models: explicit penalties

Ridge and Lasso are the canonical, fully analyzable cases — closed-form shrinkage for ridge, soft-thresholding and sparsity for Lasso, with the ball-vs-diamond geometry in Geometry of ML and the priors above. Everything else on this page is a generalization of this picture to models where the penalty cannot be written down so cleanly.

Gradient descent: early stopping as implicit \(\ell_2\)

Running GD from \(\theta_0=0\) on least squares fits fast directions (large singular values of \(X\)) first and slow directions later. Stopping at step \(t\) therefore leaves low-signal directions near zero — the same directional shrinkage pattern as ridge, with \(\lambda\sim 1/(\eta t)\) as the rough correspondence. Early stopping is a regularizer you get for free from the optimizer; its "hyperparameter" is the validation-based stopping time. See Gradient Descent.

Trees: structural constraints

Trees have no weights to penalize, so capacity is constrained structurally: maximum depth, minimum samples per leaf/split, and cost-complexity pruning

\[ R_\alpha(T)=R(T)+\alpha\,\vert T\vert, \]

which is the penalized-objective template with \(R(\theta)\) replaced by leaf count \(\vert T\vert\). Depth limits act before fitting (pre-pruning), cost-complexity acts after (post-pruning). See Decision Tree.

Bagging: regularization by averaging

Bagging changes no objective and no single model — it reduces variance by averaging \(B\) resampled fits, leaving bias untouched (the \(\rho\)-limited variance formula is in Bias–Variance). It is the regularizer of choice when the base learner is deliberately overfit (deep trees), which is exactly the random-forest recipe. Boosting's counterpart knob is shrinkage: scaling each round by a learning rate \(\nu<1\) plus early stopping on rounds. See Ensemble Methods.

Deep networks: decay, dropout, augmentation, smoothing

  • Weight decay subtracts \(\eta\lambda\theta\) in each update. With plain SGD this equals an \(\ell_2\) loss penalty; with adaptive optimizers it does not, which is why AdamW decouples decay from the gradient step (see Optimization Methods).
  • Dropout masks activations with probability \(p\) during training and rescales at test time. Two readings: an implicit ensemble of exponentially many subnetworks, and noise injection that penalizes co-adapted features. Not equivalent to \(\ell_2\) except in special linear cases.
  • Data augmentation is a prior over invariances expressed as data: if labels are invariant to a transform group (flips, crops, shifts), sampling that group constrains the learned function to respect it — capacity control in function space rather than parameter space. See CNN.
  • Label smoothing replaces one-hot targets with \((1-\epsilon)\,y+\epsilon/K\), capping achievable confidence and shrinking logit magnitudes — an output-space analogue of weight shrinkage.

See Neural Networks.


When to Prefer Which

Situation Preferred regularizer Why
Many features, few believed relevant \(\ell_1\) / Elastic Net Sparsity doubles as selection
Correlated features, all mildly useful \(\ell_2\) Stable shrinkage, no arbitrary dropping
Single tree overfits Depth / min-samples limits, then bagging Structural first, averaging second
Any high-variance learner, compute available Bagging Variance reduction without retuning the learner
Deep net, limited data Augmentation first, then decay + dropout Invariance priors are the cheapest capacity cut
Long training runs Early stopping Free, needs only a clean validation split
Overconfident classifier Label smoothing Directly targets the confidence pathology

Rule of thumb: prefer the regularizer whose assumption matches your prior knowledge — sparsity, smoothness, invariance, or "the model is simply too flexible" — rather than stacking all of them by default.


Validation Rule

Regularization strength is selected inside the training/validation procedure. Feature scaling, imputation, and selection must be fitted separately inside every fold; fitting them once on all data leaks validation information. Selection workflow: Model Selection Guide.


Connections