Configure the escape network
Escape conditions
Protect early layers
Keep the minimum gradient above the target threshold.
OpenReach target accuracy
Improve validation performance without unstable training.
OpenStay efficient
Respect the parameter and intervention limits.
OpenAvoid explosions
Prevent gradient norms from exceeding the safe range.
OpenLayer health inspector
Load a comparison preset
Presets change the current controls and refresh the expected comparison chart. Start training to test the selected configuration.
Escape result
Achievements
Experiment history
| Date | Challenge | Architecture | Activation | Health | Accuracy | Score | Outcome |
|---|---|---|---|---|---|---|---|
| No completed runs yet. | |||||||
Why the choices matter
Activation saturation
Sigmoid and tanh can compress large inputs into flat regions. Their tiny derivatives weaken gradients across many layers.
Initialization scale
Xavier, He, and LeCun methods match weight variance to network structure. Better scaling preserves signal strength during forward and backward passes.
Residual paths
Shortcut connections create shorter routes for information and gradients. They make very deep networks easier to optimize.
Recurrent memory
LSTM and GRU gates preserve useful state across longer sequences. Vanilla recurrent networks lose distant information more easily.