Dropout as a Regularizer of Interaction Effects
This provides a theoretical explanation for Dropout's effectiveness in preventing overfitting to spurious high-order interactions in neural networks, which is incremental but clarifies its role compared to other regularizers.
The paper tackles the problem of understanding Dropout's regularization mechanism by analyzing it through the lens of interactions, showing that Dropout regularizes against higher-order interactions with a decay rate of (1-p)^k, and proves this both analytically and empirically.
We examine Dropout through the perspective of interactions. This view provides a symmetry to explain Dropout: given $N$ variables, there are ${N \choose k}$ possible sets of $k$ variables to form an interaction (i.e. $\mathcal{O}(N^k)$); conversely, the probability an interaction of $k$ variables survives Dropout at rate $p$ is $(1-p)^k$ (decaying with $k$). These rates effectively cancel, and so Dropout regularizes against higher-order interactions. We prove this perspective analytically and empirically. This perspective of Dropout as a regularizer against interaction effects has several practical implications: (1) higher Dropout rates should be used when we need stronger regularization against spurious high-order interactions, (2) caution should be exercised when interpreting Dropout-based explanations and uncertainty measures, and (3) networks trained with Input Dropout are biased estimators. We also compare Dropout to other regularizers and find that it is difficult to obtain the same selective pressure against high-order interactions.