VB-Mitigator
Find the bias your accuracy score hides — and fix it.
A model can look great on average and still fail badly for specific groups of people. VB-Mitigator is an open toolkit that makes those hidden failures visible and gives you proven methods to reduce them — through a friendly point-and-click app, no deep-learning expertise required.
The one-minute version
Train an image model, and VB-Mitigator shows you not just “91% accurate” but “…yet only 46% accurate on one subgroup.” Then it lets you switch on a bias-mitigation method and watch that gap close — all from a web app you launch with one command.
Why it matters
Imagine a model that predicts a person’s gender from a photo. On paper it scores well. But if the training data accidentally links gender with skin tone, the model may quietly learn that shortcut — and then misread people whose appearance doesn’t match the shortcut. Averages hide this; subgroups reveal it.
VB-Mitigator was built to make that reality easy to see and easy to act on:
👁️ See the bias. Per-subgroup accuracy heatmaps, worst-group callouts, and real example images of who the model gets wrong.
🛠️ Reduce the bias. 16 published mitigation methods, switchable with a click.
⚖️ Compare fairly. Judge methods on their worst group, not just the average.
🖥️ Use it without code. A web dashboard launches, monitors and compares runs live.
A real example on faces: a plain model leaves one subgroup far behind (red); a mitigation method lifts it back up (green).
Who is this for?
Researchers comparing bias-mitigation methods on standard benchmarks.
Practitioners & students who want to understand model fairness hands-on.
Anyone who has heard “AI can be biased” and wants to see and address it concretely.
Start here
🚀 Get started — install it and run your first experiment from the app.
💡 What is visual bias? — a plain-English tour of the problem.
🖥️ Using the app — launch, watch training live, and compare runs.
📊 Reading the results — heatmaps, worst-group, the fairness gap.
📚 Catalog — the datasets, methods, models and metrics in plain English.
Supported by the EU Horizon Europe projects MAMMOth (GA 101070285) and ELIAS (GA 101120237). Released under the MIT License.