Model Stabilization
Stabilize Your Model
Unstable models will never reach their full potential. By using our patented weights analysis we can identify if your model is unstable.
Unstable Models
An UNSTABLE model will NOT generalize and is characterized by:
-
Fewer than 30% of weights change in later epochs
-
NaN errors
-
Erratic loss curve
Fixes
New Backbone
Upon analysis we can tell you if your dataset is benefiting from transfer learning. If not, we can build you a “Backbone” model to use as your base model.
Error & Loss Curve Corrections
If your error-loss graph over the training run is erratic and doesn't follow an ideal curve, this is indicative of a unstable model that will never reach its potential. We will stabilize your model and put you on the path to generalization.
Break Out of the Accuracy Plateau
If your accuracy/F1 score has reached a plateau, we can help stabilize your model to attain increased accuracy.
