Dropout addresses co-adaptation during training by randomly disabling neurons with probability p and keeping activation with q=1-p. Inputs are masked and scaled by 1/q to preserve expected magnitude. In backprop, the same mask scales gradients; in inference, masking is disabled.

Implementation adds a Dropout layer class (CNeuronDropoutOCL) with a training flag at base-neuron level. Forward pass builds a mask, transfers it to GPU, and performs element-wise multiplication via an OpenCL kernel using vector types (double4). Backward pass reuses the same kernel to apply the mask to gradients.

Persistence stores only dropout probability; masks are regenerated per training step. Base classes are updated to register the new layer type and kernel wiring.

👉 Read | AlgoBook | @mql5dev

#MQL5 #MT5 #Dropout