Dropout: A Simple Way to Prevent Neural Networks from Overfitting (Dropout)

Implements probabilistic neuron dropout by randomly deactivating a fraction of neurons (typically 0.5) during each forward-backward training pass, forcing the network to learn redundant representations across different neuron subsets. The mechanism works by applying element-wise multiplication of activations by Bernoulli random variables sampled independently per training iteration, effectively creating an ensemble of thinned networks that share weights. At test time, activations are scaled by the dropout probability to maintain expected values, or inverted dropout rescales during training instead.

Extends basic dropout with learned or scheduled dropout rates that vary across layers and training phases, allowing different network depths to use different dropout probabilities (e.g., higher rates for early layers, lower for final classification layers). Implementation uses layer-specific dropout parameters that can be tuned via validation performance or learned through auxiliary loss terms, enabling automatic discovery of optimal regularization strength per layer without manual grid search.

Applies dropout to recurrent neural networks (RNNs, LSTMs, GRUs) by using the same dropout mask across all timesteps within a sequence, rather than sampling independent masks per timestep. This preserves temporal dependencies while preventing co-adaptation of recurrent connections. Implementation maintains a fixed Bernoulli mask for the entire sequence length, then applies it consistently to hidden state transitions, enabling effective regularization without disrupting the recurrent information flow that would occur with per-timestep dropout.

Applies dropout to convolutional networks by dropping entire feature maps (channels) rather than individual activations, preserving spatial structure within feature maps while preventing co-adaptation across channels. Implementation samples a single Bernoulli mask per channel and applies it uniformly across all spatial locations (height × width), maintaining spatial coherence in learned features. This is particularly effective for image data where spatial relationships are semantically meaningful.

Repurposes dropout as a Bayesian approximation by performing multiple stochastic forward passes at test time with dropout enabled, treating each pass as a sample from the posterior distribution over model weights. Implementation runs the same input through the network 10-100 times with different random dropout masks, collecting predictions from each pass to estimate prediction uncertainty via variance across samples. This provides calibrated confidence estimates without retraining or architectural changes, approximating Bayesian inference through repeated stochastic sampling.

Leverages the implicit ensemble created by dropout during training by averaging predictions from multiple forward passes at test time, where each pass uses a different random dropout mask. Unlike Monte Carlo dropout which uses dropout for uncertainty estimation, this capability focuses on pure ensemble averaging for improved accuracy. Implementation runs inference 5-20 times with dropout enabled and averages the output logits or probabilities, effectively combining predictions from different thinned network configurations to reduce variance and improve generalization.