Batch Size Explorer

See how the amount of data per gradient update affects the descent path.

Epoch 0
Step: 0
Scenario

Convex loss surface from linear regression. All methods converge, but watch the paths differ.

Learning Rate 0.010
Mini-batch Size 16
Animation Speed 1x
Batch
0.00
0 steps
Running
Mini-batch
0.00
0 steps
Running
SGD
0.00
0 steps
Running
Understanding the Trade-offs

Batch GD uses all 100 points per step (accurate but slow). SGD uses 1 point (fast but noisy). Mini-batch balances both.

Dataset & Fit Lines
Batch (all 100)
Mini (16)
SGD (1)
Why Noise Helps SGD's noisy gradient can escape local minima where batch GD gets stuck. Try the "Local Trap" scenario!