Home▸Data Science▸Feature Scaling Lab 2D: Gradient Descent Race Track

Feature Scaling Lab 2D: Gradient Descent Race Track

2D top-down race track comparing StandardScaler, MinMaxScaler and RobustScaler: pan and zoom across four scatter lanes as the same gradient-descent update runs on raw vs. scaled features, with a live outlier stress test.

Data Science2DModerate60 FPS📱 Mobile-adapted⇄ 3D version

Details

CategoryData Science

RenderingCanvas 2D

DifficultyModerate

Target FPS60

ControlsData-points, outlier-strength and shared-learning-rate sliders reshape the dataset and retrain live; New dataset & retrain regenerates; Reset View restores default pan/zoom; click a lane box to focus its stats. Drag the canvas to pan across lanes and scroll/wheel to zoom into a track.

feature scalinggradient descentstandardizationnormalizationstandardscalerminmaxscalerrobustscaler

Related simulations

Dimensionality Reduction: PCA, t-SNE & UMAPWatch real PCA (covariance eigendecomposition), a live t-SNE-style optimisation, and a…

Regression Visualizer: Linear, Polynomial & RegularizationFit noisy data live with real least-squares regression — slide the polynomial degree to see…

Time Series Decomposition: Trend, Seasonality & ResidualsDecompose a synthetic time series into trend, seasonality and residuals in real time —… ⊞ All 156 in Data Science →

This is the top-down companion to the 3D Feature Scaling Lab: the same synthetic dataset (income and age predicting an outcome) runs through the same four lanes — raw and unscaled, StandardScaler (z-score), MinMaxScaler (0–1 range) and RobustScaler (median/IQR) — each fitting an identical linear model with plain gradient descent at one shared learning rate. Instead of orbiting a 3D scene, you pan and zoom a flat race track: each lane's scatter shows the two features plotted against each other to scale, and a vertical "loss tower" behind it rises with log-loss exactly as the raw lane's tower shoots up and flashes red while the scaled lanes stay low and steady. Inject a single extreme outlier to watch StandardScaler and especially MinMaxScaler wobble while RobustScaler's median/IQR math keeps it unmoved.

⚙ Under the hood

Top-down 2D race track comparing four lanes of the same dataset — raw, StandardScaler, MinMaxScaler and RobustScaler — as each fits a linear model with the identical batch gradient-descent update and one shared learning rate. Watch the raw lane's loss tower shoot up and flash red while scaled lanes stay low, then inject an outlier to see MinMaxScaler wobble while RobustScaler holds steady. Drag to pan across lanes and scroll to zoom into any one track.

feature scalinggradient descentstandardizationnormalizationstandardscalerminmaxscalerrobustscaler

2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install