Calculator inputs
Choose one analysis mode. Relevant input fields will appear below.
Formula used
How to use
Select the calculation mode matching your network. Enter valid dimensions and dropout settings. Then press the calculate button.
Review expected retained and dropped values. Inspect the chart for visual comparison. Export results when needed.
Use Monte Carlo mode for uncertainty analysis. Increase simulations for a steadier estimate. Avoid excessive rates that cause underfitting.
Example data
| Total neurons | Dropout rate | Expected active | Expected dropped |
|---|---|---|---|
| 100 | 0.20 | 80 | 20 |
| 256 | 0.25 | 192 | 64 |
| 512 | 0.50 | 256 | 256 |
| 1,000 | 0.30 | 700 | 300 |
Calculation history
Saved results remain in this browser until cleared.
| Date | Mode | Summary | Action |
|---|
Dropout guidance
Dropout randomly suppresses units during training. This reduces co-adaptation between learned features. It often improves generalization.
Dense layers commonly tolerate moderate dropout. Convolutional layers may benefit from spatial dropout. Recurrent models require careful masking.
Inference usually disables random masking. Inverted dropout already preserves expected activation scale. Validation performance should guide selection.
Frequently asked questions
What is dropout in machine learning?
Dropout is a regularization method that randomly disables units during training. It discourages dependence on specific neurons. This can improve generalization.
How is the dropout rate calculated?
The dropout rate is the removed-unit probability. A rate of 0.20 drops twenty percent. The remaining probability equals 0.80.
What is keep probability?
Keep probability is one minus dropout rate. It measures the chance of retention. Higher values preserve more activations.
Why is inverted dropout used?
Inverted dropout scales retained activations during training. Their expected magnitude remains stable. Inference needs no additional scaling.
Is dropout active during inference?
Standard inference usually disables random dropout masks. Monte Carlo dropout intentionally keeps them active. That approach estimates predictive uncertainty.
What dropout rate should I use?
Useful values often range from 0.10 to 0.50. The best setting depends on architecture. Validate several rates empirically.
Can dropout be applied to convolutional layers?
Yes, but spatial dropout may work better. It removes entire feature maps together. This respects local spatial correlation.
What is spatial dropout?
Spatial dropout suppresses complete channels or maps. Standard dropout removes individual activation elements. Spatial masking can regularize convolutional features.
Can dropout cause underfitting?
Yes, excessive dropout reduces effective model capacity. Training loss may remain high. Lower the rate when learning stalls.
Does dropout reduce model file size?
Usually it does not remove stored parameters. Dropout affects temporary training activations. Parameter tensors generally remain unchanged.
What is Monte Carlo dropout?
Monte Carlo dropout repeats stochastic forward passes. Variation across outputs estimates uncertainty. More runs stabilize the estimate.