Sourced from whest-starterkit @ aaa3882.

Use WhestBench Explorer

← Documentation

🎯 When to use this page

When you want visual intuition about network behavior and where your estimator's error concentrates. The Explorer is optional: it is not the submission interface, and your leaderboard score never depends on it.

🚀 Open it

The Explorer is a separate, hosted React app:

Open the hosted URL, generate an MLP, propagate inputs, and inspect activations layer-by-layer. There's nothing to install.

The Explorer used to ship inside whestbench as a whest visualizer subcommand. As of whestbench commit 28c203f (May 2026), it lives in its own repo with auto-deploy to GitHub Pages — whest visualizer no longer exists.

✅ Expected outcome

An interactive view of network structure, layer behavior, and estimator-vs-ground-truth comparisons.

Suggested workflow

1. Start with small width/depth.
2. Vary the seed to inspect how structure changes.
3. Compare estimator behavior across layers.
4. Locate where errors concentrate.
5. Convert observations into Python estimator heuristics, then verify with:

   uv run whest run --estimator estimator.py --runner local

The Explorer is for intuition — it is not a scoring oracle. Official scoring still comes from whest run.

Interpreting the visualization

The Explorer shows neuron activations across layers:

• Rows: layers (top = first layer, bottom = output)
• Columns: neurons within each layer
• Color intensity: mean activation value

Patterns to look for:

• Error grows at deep layers: your method loses accuracy as correlations accumulate through layers.
• Sudden drops to zero: ReLU is killing neuron groups — your variance estimates may be too narrow.
• Uniform predictions: your estimator may not be exploiting the weight structure.

➡️ Next step

• Validate, Run, and Package
• Problem Setup