Low-dimensional intrinsic dimension reveals a phase transition in gradient-based learning of deep neural networks

Deep neural networks complete a feature extraction task by propagating the inputs through multiple modules. However, how the representations evolve with the gradient-based optimization remains unknown. Here we leverage the intrinsic dimension of the representations to study the learning dynamics and find that the training process undergoes a phase transition from expansion to compression under disparate training regimes. Surprisingly, this phenomenon is ubiquitous across a wide variety of model architectures, optimizers, and data sets. We demonstrate that the variation in the intrinsic dimension is consistent with the complexity of the learned hypothesis, which can be quantitatively assessed by the critical sample ratio that is rooted in adversarial robustness. Meanwhile, we mathematically show that this phenomenon can be analyzed in terms of the mutable correlation between neurons. Although the evoked activities obey a power-law decaying rule in biological circuits, we identify that the power-law exponent of the representations in deep neural networks predicted adversarial robustness well only at the end of the training but not during the training process. These results together suggest that deep neural networks are prone to producing robust representations by adaptively eliminating or retaining redundancies. The code is publicly available at https://github.com/cltan023/learning2022.