Choosing a suitable loss function is essential when learning by empirical risk minimisation. In many practical cases, the datasets used for training a classifier may contain incorrect labels, which prompts the interest for using loss functions that are inherently robust to label noise. In this paper, we study the Fisher-Rao loss function, which emerges from the Fisher-Rao distance in the statistical manifold of discrete distributions. We derive an upper bound for the performance degradation in the presence of label noise, and analyse the learning speed of this loss. Comparing with other commonly used losses, we argue that the Fisher-Rao loss provides a natural trade-off between robustness and training dynamics. Numerical experiments with synthetic and MNIST datasets illustrate this performance.

A full-rank lattice in the Euclidean space is a discrete set formed by all integer linear combinations of a basis. Given a probability distribution on ${\mathbb{R}}^n$, two operations can be induced by considering the quotient of the space by such a lattice: wrapping and quantization. For a lattice $\Lambda$, and a fundamental domain $\mathcal{D}$, which tiles ${\mathbb{R}}^n$ through $\Lambda$, the wrapped distribution over the quotient is obtained by summing the density over each coset, while the quantized distribution over the lattice is defined by integrating over each fundamental domain translation. These operations define wrapped and quantized random variables over $\mathcal{D}$ and $\Lambda$, respectively, which sum up to the original random variable. We investigate information-theoretic properties of this decomposition, such as entropy, mutual information and the Fisher information matrix, and show that it naturally generalizes to the more abstract context of locally compact topological groups.

In Table 1, row "Negative multinomial", column "Support", "m" should be corrected as "n". 2. In Table 1, rows "Categorical" and "Negative multinomial", column "Parameters", the first comma next to " ∈ " has been wrongly placed. Moreover, a comma should be placed between "]0, 1[ n-1 " and " i". 3. In Table 1, row "Negative multinomial", column "References", it should be "[12,19,48]" instead of "[12,19,46]". 4. In Table 2, row "Weibull", column "Fisher-Rao distance", the roles of β 1 and β 2 should be inverted. 5. In Table 2, row "Power function", α in all columns should be corrected as β. Moreover, the column "Fisher matrix", "α" should be replaced by "β".