Learn AIDiagnose a training run.
This playground now belongs to a full lesson and hands-on project.
Open the six-stage lesson · Project and independent transfer assessment. Learner pilots remain pending.
This 13-parameter network fits the supplied curve y = 0.5x + 0.35 sin(3x). Gradients use 32 training inputs; 32 different interleaved inputs monitor validation. Both come from the same known synthetic function. Repeated validation is development evidence, not an untouched final test. Capture matched runs below before navigating away.
Experiment: change one cause
Changing a training control restarts from the same initialization and clears optimizer moments. Save a comparison first. No automatic training or animation starts.
Choose a failure case, predict its signature, then train.
0 / 1,000accepted updates
6.089e-2training mean squared error
6.257e-2validation mean squared error
0 / 128hidden outputs with |h| > 0.99
Computed gradient norm: 3.534e-1 · gradient norm sent to optimizer: 3.534e-1. A nonzero computed gradient with zero applied gradient is evidence of the simulated blocked-gradient bug. Small gradients with many saturated units suggest inspecting scale.
Inspect the actual calculation
Scaled input: 0.50 · prediction: 0.408154 · target: 0.599123 · error: -0.190969.
| Parameter | Value | Computed gradient | Applied gradient | Proposed change |
|---|---|---|---|---|
| w₁[0] | 0.400000 | -0.039147 | -0.039147 | 3.9147e-3 |
Playground values are temporary. Capture a finished run to save project evidence locally; captures and reports are included in My Work backups. Four evidence slots are available per project.