Noise is traditionally treated as an adversary in signal processing and machine learning. Extensive effort has been devoted to maximising Signal-to-noise ratio (SNR), modelling noise distributions, and denoising the signal to recover underlying structure. In statistical learning, this reduction of observational noise improves estimation accuracy. Yet in modern deep learning, noise plays a paradoxically constructive role. When introduced in a controlled manner, stochastic perturbations improve robustness, mitigate overfitting to spurious correlations, and promote invariant representation learning.
In this talk, I trace the formalisation of dropout-induced stochasticity as a learning mechanism: from stochastic regularisation, to latent-variable modelling in Variational Autoencoders (VAEs), to information-theoretic compression in Variational Information Bottleneck (VIB). I discuss how structured noise reshapes the optimisation landscape and acts as an implicit bias toward approximately minimal sufficient representations.
I conclude with a practical demonstration in digital image forensics, showing how noise-induced learning improves detection of manipulated and deepfake imagery in the generative AI era.
