Optimization

Weight Decay & AdamW: Decoupled Regularization

Why shrinking weights is not the same as adding an L2 penalty inside Adam, and how AdamW restores the intended regularization behavior.

status: reviewimportance: importantdifficulty 3/5math: undergraduateread: 13mdemo planned

Back to Optimization Next: Scaling Laws & Emergent Abilities

Concept Structure

Weight Decay & AdamW: Decoupled Regularization

01Intuition

Start with the picture, metaphor, or geometric mechanism.

02Math

Make the objects explicit and connect them with notation.

03Code

Mirror the equations with runnable implementation details.

04Interactive Demo

Manipulate the mechanism and watch the idea respond.

2prerequisites

1next concepts

2related links

Learning map

Weight Decay & AdamW: Decoupled Regularization

BeforeAdam OptimizerNow3/4 sections readyTryUse the demo notes to predict the mechanism before moving on.NextScaling Laws & Emergent Abilities

Object flow

3/4 sections readyAsk about this Research room

ConceptWeight Decay & AdamW: Decoupled RegularizationOptimization EquationWeight Decay & AdamW: Decoupled Regularization equation 1Exact equation object CodeWeight Decay & AdamW: Decoupled Regularization code witness 1Exact code witness ClaimWeight Decay & AdamW: Decoupled Regularization central claimMechanism claim EquationWeight Decay & AdamW: Decoupled Regularization math objectsEquation and notation CodeWeight Decay & AdamW: Decoupled Regularization code witnessRunnable code object

ConceptWeight Decay & AdamW: Decoupled RegularizationOptimization

Local snapshot ready

concept:optimization/weight-decay-adamw

Codewitness nearby Predictbefore reveal Roomobject handoff

Conceptual Bridge

What should feel connected as you move through this page.

Carry inAdam Optimizer

Bring the mental model from Adam Optimizer; this page will reuse it instead of restarting from zero.

Work hereWeight Decay & AdamW: Decoupled Regularization

Why shrinking weights is not the same as adding an L2 penalty inside Adam, and how AdamW restores the intended regularization behavior.

Carry outScaling Laws & Emergent Abilities

The next edge should feel earned: use the demo prediction here before following Scaling Laws & Emergent Abilities.

Test the linkUse the demo notes to predict the mechanism before moving on.Then continue to Scaling Laws & Emergent Abilities

01IntuitionStart with the picture, metaphor, or geometric mechanism.02MathMake the objects explicit and connect them with notation.03CodeMirror the equations with runnable implementation details.04Interactive DemoManipulate the mechanism and watch the idea respond.

Intuition

Build the mental picture first so the rest of the page has something to attach to.

Section prompt

Regularization often starts with a simple preference: all else equal, prefer smaller weights.

There are two closely related ways to express that idea:

add an $L_2$ penalty to the loss,
directly shrink parameters a little bit every step.

For plain SGD, those are effectively the same update. That is why many people casually treat "L2 regularization" and "weight decay" as synonyms.

For Adam, they are not the same. Adam rescales coordinates using running estimates of gradient magnitude, so an $L_2$ term added to the gradient gets rescaled too. That means different parameters can experience very different effective regularization strengths.

AdamW fixes this by decoupling weight decay from the adaptive gradient step. First do the Adam update. Then shrink the parameters directly.

Math

Translate the story into symbols, assumptions, and a derivation you can inspect.

Section prompt

Equation 1L_{reg}(\theta) = L(\theta) + \frac{\lambda}{2} \lVert \theta \rVert_2^2,Equation 2\nabla L_{reg}(\theta) = \nabla L(\theta) + \lambda \theta.

Let $L(\theta)$ be the task loss and $\lambda > 0$ the regularization strength.

With an $L_2$ penalty, the regularized objective is

L_{reg}(\theta) = L(\theta) + \frac{\lambda}{2} \lVert \theta \rVert_2^2,

so the gradient becomes

\nabla L_{reg}(\theta) = \nabla L(\theta) + \lambda \theta.

For SGD, this leads to

\theta_{t+1} = \theta_t - \eta \nabla L(\theta_t) - \eta \lambda \theta_t = (1 - \eta \lambda)\theta_t - \eta \nabla L(\theta_t),

which is exactly a weight-decay step.

For Adam, the adaptive preconditioner changes things. AdamW writes the update as

m_t = \beta_1 m_{t-1} + (1-\beta_1) g_t,

v_t = \beta_2 v_{t-1} + (1-\beta_2) g_t^2,

\theta_{t+1} = \theta_t - \eta \frac{\hat m_t}{\sqrt{\hat v_t} + \epsilon} - \eta \lambda \theta_t.

The key point is that the shrinkage term $-\eta \lambda \theta_t$ is not divided by $\sqrt{\hat v_t} + \epsilon$ . Regularization stays regularization instead of getting entangled with Adam's coordinatewise scaling.

Code

Keep the implementation aligned with the notation so the algorithm is legible.

Section prompt

Code witness 1import numpy as np theta = np.array([1.0, 1.0]) g = np.array([0.1, 0.1]) vhat = np.array([1e-...python

import numpy as np

theta = np.array([1.0, 1.0])
g = np.array([0.1, 0.1])
vhat = np.array([1e-4, 1.0])  # Adam sees very different curvature/noise
lr = 1e-2
wd = 0.1
eps = 1e-8

adam_with_l2 = theta - lr * ((g + wd * theta) / (np.sqrt(vhat) + eps))
adamw = theta - lr * (g / (np.sqrt(vhat) + eps)) - lr * wd * theta

print("Adam + L2 :", np.round(adam_with_l2, 4))
print("AdamW     :", np.round(adamw, 4))

In the first coordinate, the effective shrinkage under "Adam + L2" becomes much larger because Adam divides by a tiny sqrt(vhat). AdamW avoids that distortion.

Interactive Demo

Use direct manipulation to connect the explanation to a moving system.

Section prompt

No interactive demo yet for this concept. A useful next step would be a two-coordinate optimizer sandbox showing how Adam + L2 and AdamW diverge when one coordinate has much smaller running variance.

No live visualization is registered for this concept yet.

The page still supports explanatory demo notes above; when a viz.tsx exists, it will render here without changing the route.

After The First Pass

Turn the concept into an inspected object.

Once the invariant is visible in the intuition, math, code, and demo, use these panels to inspect the mechanism visually, check source support, practice the idea, and attach a grounded research question.

Mechanism Storyboard

See the idea move before the page explains it

Why shrinking weights is not the same as adding an L2 penalty inside Adam, and how AdamW restores the intended regularization behavior.

Demo notes open01 / Intuition

Prediction lens

Start with the picture, metaphor, or geometric mechanism.

Commit first

Before reading further, choose the kind of change Weight Decay & AdamW: Decoupled Regularization should make visible.

Visual Inquiry

Make the image answer a mathematical question

Why shrinking weights is not the same as adding an L2 penalty inside Adam, and how AdamW restores the intended regularization behavior.

3/4 stages readyDemo notes connected

Visual cueWhich visible object should carry the first intuition?

Inspection depth2/4

Prediction

Which visible object should carry the first intuition?

Commit first

Pick the cue that should make Weight Decay & AdamW: Decoupled Regularization easier to reason about before the page gives the answer.

Claim Review

Why shrinking weights is not the same as adding an L2 penalty inside Adam, and how AdamW restores the intended regularization behavior.

StatusSubstantive claim review pending

Source IDs and witness objects are attached for review; they are not proof by themselves.

SourcesNo references

Add source metadata before claiming support.

Witnesses3 local objects

Use equation, code, and demo objects to check whether the source support is operational.

Source support candidates

No structured source note is attached yet.

Mechanism witnesses

Equation 1

L_{reg}(\theta) = L(\theta) + \frac{\lambda}{2} \lVert \theta \rVert_2^2,

Equation 2

\nabla L_{reg}(\theta) = \nabla L(\theta) + \lambda \theta.

Code witness 1import numpy as np theta = np.array([1.0, 1.0]) g = np.array([0.1, 0.1]) vhat = np.array([1e-...

Practice Loop

Try the idea before it explains itself

Why shrinking weights is not the same as adding an L2 penalty inside Adam, and how AdamW restores the intended regularization behavior.

Readiness0/3 checks ready

Predict

Before touching the demo, predict one visible change that should happen in Weight Decay & AdamW: Decoupled Regularization.

Hint 1

Reveal when your model needs a nudge.

Hint 2

Reveal when your model needs a nudge.

Hint 3

Reveal when your model needs a nudge.

Your answer canvas

Local checks

Claim

A concrete answer is on the canvas.

Mechanism

The answer names why the claim should hold.

Bridge

It touches the page context or a neighboring idea.

Misconception check

Object research drawerClose

ConceptWeight Decay & AdamW: Decoupled RegularizationOptimization

Code witness comparisonWeight Decay & AdamW: Decoupled Regularization code witness 1theta = np.array([1.0, 1.0])Prediction before revealWeight Decay & AdamW: Decoupled Regularization predictionUse the demo notes to predict the mechanism before moving on.

Grounded room questionWhat is the smallest example that makes Weight Decay & AdamW: Decoupled Regularization click without losing the math?Local snapshot ready

Research Room

Attach the question to an exact object

Pick the concept, equation, source, code witness, claim, misconception, or demo state before asking for help. The handoff stays grounded to that object.

Next local actionNo local draft saved yet

Open the draft below to save one note and next action in this browser.

conceptOptimization

Weight Decay & AdamW: Decoupled Regularization

Anchored question

What is the smallest example that makes Weight Decay & AdamW: Decoupled Regularization click without losing the math?

Local action draftNo local draft saved yetExpand only when ready to capture one local next action

Local action draft

This draft stays locally in this browser for concept:optimization/weight-decay-adamw.

Draft noteNext action

No local draft saved.

Evidence to inspect

Definition, prerequisite, and contrast concept links
The equation or code witness that makes the concept operational
One demo state that shows the invariant instead of a slogan

What would resolve this

The learner can state the mechanism in their own words
The learner can name the prerequisite that would repair confusion
The learner can predict how the mechanism changes under one perturbation

Grounded AI handoff

I am working in Continuous Function's research reading room. Object: concept - Weight Decay & AdamW: Decoupled Regularization Object key: concept:optimization/weight-decay-adamw Context: Optimization Anchor id: concept/concept-notebook/optimization/weight-decay-adamw Open question: What is the smallest example that makes Weight Decay & AdamW: Decoupled Regularization click without losing the math? Evidence to inspect: - Definition, prerequisite, and contrast concept links - The equation or code witness that makes the concept operational - One demo state that shows the invariant instead of a slogan What would resolve this: - The learner can state the mechanism in their own words - The learner can name the prerequisite that would repair confusion - The learner can predict how the mechanism changes under one perturbation Answer as a careful research tutor: stay source-grounded, separate verified evidence from assumptions, name the relevant math objects, and end with one next action.

Open source object

concept/concept-notebook/optimization/weight-decay-adamw
concept:optimization/weight-decay-adamw

Learning Map

Before / Now / Try / Next

BeforeAdam Optimizer

NowIntuition → Math → Code → Demo

TryUse the demo notes to predict the mechanism before moving on.

NextScaling Laws & Emergent Abilities

Intuitionready
Mathready
Codeready
Interactive Demoplanned

Object Companion

Ask beside the selected object

Why shrinking weights is not the same as adding an L2 penalty inside Adam, and how AdamW restores the intended regularization behavior.

Your question

GoalComfortStyleStuck on

Context prompt

You are my AI learning companion for Continuous Function. Current context: Optimization concept. Learning surface: Weight Decay & AdamW: Decoupled Regularization. What this page says: Why shrinking weights is not the same as adding an L2 penalty inside Adam, and how AdamW restores the intended regularization behavior. Current section: Intuition, math, code, and interactive demo. Suggested next step: Use the demo notes to predict the mechanism before moving on.. Learner goal: Understand the idea. Learner comfort level: New to this. Preferred explanation style: Visual first. Task: Explain the central idea in plain language, then restate it with the exact math objects from the page. Answer in a way that helps me learn: ask one clarifying question only if needed, use intuition before notation, and end with one thing I should try on the page.

Domain

Optimization

optimizationregularizationadamwweight-decay

Prerequisites

Adam Optimizer Maximum Likelihood

Leads To

Scaling Laws & Emergent Abilities

SGD & Momentum: The Workhorses of Optimization Label Smoothing & Soft Targets

Within this domain