Finite-time Lyapunov exponents of deep neural networks

We compute how small input perturbations affect the output of deep neural networks, exploring an analogy between deep networks and dynamical systems, where the growth or decay of local perturbations is characterised by finite-time Lyapunov exponents. We show that the maximal exponent forms geometrical structures in input space, akin to coherent structures in dynamical systems. Ridges of large positive exponents divide input space into different regions that the network associates with different classes. These ridges visualise the geometry that deep networks construct in input space, shedding light on the fundamental mechanisms underlying their learning capabilities.

Looking at the posterior: on the origin of uncertainty in neural-network classification

Bayesian inference can quantify uncertainty in the predictions of neural networks using posterior distributions for model parameters and network output. By looking at these posterior distributions, one can separate the origin of uncertainty into aleatoric and epistemic. We use the joint distribution of predictive uncertainty and epistemic uncertainty to quantify how this interpretation of uncertainty depends upon model architecture, dataset complexity, and data distributional shifts in image classification tasks. We conclude that the origin of uncertainty is subjective to each neural network and that the quantification of the induced uncertainty from data distributional shifts depends on the complexity of the underlying dataset. Furthermore, we show that the joint distribution of predictive and epistemic uncertainty can be used to identify data domains where the model is most accurate. To arrive at these results, we use two common posterior approximation methods, Monte-Carlo dropout and deep ensembles, for fully-connected, convolutional and attention-based neural networks.