Solvers
Overview
Compare Vyntri adaptation and shrinkage options for fast, analytic vision intelligence.
Vyntri uses analytic ridge classification with optional feature adaptation — zero gradient descent, zero epochs, zero hyperparameter sweeps. Advanced adaptation and shrinkage options available for research use.
Quick Comparison
| FK (Default) | SLCE | Ledoit-Wolf | |
|---|---|---|---|
| Full Name | Friedman's Discriminative Whitening | Supervised Linear Centroid-Encoder | Data-Estimated Shrinkage |
| How it works | Eigenvalue flooring + diagonal shrinkage | Centroid projection + eigenvalue flooring | Cross-validated shrinkage intensity |
| Speed | Fast | Fast | Fast |
| Continual learning? | Yes | No (requires re-fit) | No (requires re-fit) |
| Best for | General classification | Research, centroid separation | Research, data-driven shrinkage |
When To Use Which Method
| Scenario | Recommendation |
|---|---|
| General Classification | Use FK (default) — best overall performance with continual learning support. |
| Continual Learning | Use FK + diagonal shrinkage — only supported combination for update(). |
| Research / Experiments | Use SLCE or Ledoit-Wolf — advanced options with honest trade-offs. |
| Auto Backbone | Use backbone="auto" with optional candidates to control which backbones are evaluated. |
Supported Backbones
All adaptation and shrinkage methods work identically across all five supported backbones: mobilenet_v3_small, resnet18, resnet50, efficientnet_b0, and convnext_tiny.