Home▸Machine Learning & Neural Networks▸Batch Size Impact

🧪 Batch Size Impact

Interactive SGD landscape: watch how batch size controls gradient-noise magnitude (sigma over sqrt B), steering training into a sharp, low-loss minimum or a flatter, better-generalizing one.

Machine Learning & Neural Networks2DModerate60 FPS📱 Mobile-adapted

This simulation grounds the batch-size hyperparameter in the actual math of mini-batch stochastic gradient descent: a mini-batch gradient is an average of B per-sample gradients, so its noise variance shrinks as 1/B. A ball descends a real two-well loss landscape under that noise — small batches keep enough noise to escape a narrow sharp minimum and settle into a wide flat one, while large batches converge precisely into whichever minimum is nearest, even a sharp one with lower training loss but a larger train/test generalization gap. Adjust batch size, learning rate and playback speed to watch the trade-off play out step by step.

⚙ Under the hood

This simulation investigates the effect of varying batch sizes during machine learning model training. It demonstrates how these choices influence convergence and overall performance.

Batch SizeImpact

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