Searcharxiv⌕ Search

arXiv subjects

Joseba Dalmau

Publications and source records attributed to Joseba Dalmau.

17 recordsLinked to original sources
arXiv2025

Deep Sturm--Liouville: From Sample-Based to 1D Regularization with Learnable Orthogonal Basis Functions

Although Artificial Neural Networks (ANNs) have achieved remarkable success across various tasks, they still suffer from limited generalization. We hypothesize that this limitation arises from the traditional sample-based (0--dimensionnal) regularization used in ANNs. To overcome this, we introduce \textit{Deep Sturm--Liouville} (DSL), a novel function approximator that enables continuous 1D regularization along field lines in the input space by integrating the Sturm--Liouville Theorem (SLT) into the deep learning framework. DSL defines field lines traversing the input space, along which a Sturm--Liouville problem is solved to generate orthogonal basis functions, enforcing implicit regularization thanks to the desirable properties of SLT. These basis functions are linearly combined to construct the DSL approximator. Both the vector field and basis functions are parameterized by neural networks and learned jointly. We demonstrate that the DSL formulation naturally arises when solving a Rank-1 Parabolic Eigenvalue Problem. DSL is trained efficiently using stochastic gradient descent via implicit differentiation. DSL achieves competitive performance and demonstrate improved sample efficiency on diverse multivariate datasets including high-dimensional image datasets such as MNIST and CIFAR-10.

cs.LG↗
arXiv2024

Improving Out-of-Distribution Detection by Combining Existing Post-hoc Methods

Since the seminal paper of Hendrycks et al. arXiv:1610.02136, Post-hoc deep Out-of-Distribution (OOD) detection has expanded rapidly. As a result, practitioners working on safety-critical applications and seeking to improve the robustness of a neural network now have a plethora of methods to choose from. However, no method outperforms every other on every dataset arXiv:2210.07242, so the current best practice is to test all the methods on the datasets at hand. This paper shifts focus from developing new methods to effectively combining existing ones to enhance OOD detection. We propose and compare four different strategies for integrating multiple detection scores into a unified OOD detector, based on techniques such as majority vote, empirical and copulas-based Cumulative Distribution Function modeling, and multivariate quantiles based on optimal transport. We extend common OOD evaluation metrics -- like AUROC and FPR at fixed TPR rates -- to these multi-dimensional OOD detectors, allowing us to evaluate them and compare them with individual methods on extensive benchmarks. Furthermore, we propose a series of guidelines to choose what OOD detectors to combine in more realistic settings, i.e. in the absence of known OOD data, relying on principles drawn from Outlier Exposure arXiv:1812.04606. The code is available at https://github.com/paulnovello/multi-ood.

stat.ML↗
arXiv2024

Conformal Semantic Image Segmentation: Post-hoc Quantification of Predictive Uncertainty

We propose a post-hoc, computationally lightweight method to quantify predictive uncertainty in semantic image segmentation. Our approach uses conformal prediction to generate statistically valid prediction sets that are guaranteed to include the ground-truth segmentation mask at a predefined confidence level. We introduce a novel visualization technique of conformalized predictions based on heatmaps, and provide metrics to assess their empirical validity. We demonstrate the effectiveness of our approach on well-known benchmark datasets and image segmentation prediction models, and conclude with practical insights.

cs.CV↗
arXiv2024

Out-of-Distribution Detection Should Use Conformal Prediction (and Vice-versa?)

Research on Out-Of-Distribution (OOD) detection focuses mainly on building scores that efficiently distinguish OOD data from In Distribution (ID) data. On the other hand, Conformal Prediction (CP) uses non-conformity scores to construct prediction sets with probabilistic coverage guarantees. In this work, we propose to use CP to better assess the efficiency of OOD scores. Specifically, we emphasize that in standard OOD benchmark settings, evaluation metrics can be overly optimistic due to the finite sample size of the test dataset. Based on the work of (Bates et al., 2022), we define new conformal AUROC and conformal FRP@TPR95 metrics, which are corrections that provide probabilistic conservativeness guarantees on the variability of these metrics. We show the effect of these corrections on two reference OOD and anomaly detection benchmarks, OpenOOD (Yang et al., 2022) and ADBench (Han et al., 2022). We also show that the benefits of using OOD together with CP apply the other way around by using OOD scores as non-conformity scores, which results in improving upon current CP methods. One of the key messages of these contributions is that since OOD is concerned with designing scores and CP with interpreting these scores, the two fields may be inherently intertwined.

stat.ML↗
arXiv2022

Scaling limit of the heavy-tailed ballistic deposition model with $p$-sticking

Ballistic deposition is a classical model for interface growth in which unit blocks fall down vertically at random on the different sites of $\mathbb{Z}$ and stick to the interface at the first point of contact, causing it to grow. We consider an alternative version of this model in which the blocks have random heights which are i.i.d. with a heavy (right) tail, and where each block sticks to the interface at the first point of contact with probability $p$ (otherwise, it falls straight down until it lands on a block belonging to the interface). We study scaling limits of the resulting interface for the different values of $p$ and show that there is a phase transition as $p$ goes from $1$ to $0$.

math.PR↗
arXiv2020

A probabilistic proof of Perron's theorem

We present an alternative proof of Perron's theorem, which is probabilistic in nature. It rests on the representation of the Perron eigenvector as a functional of the trajectory of an auxiliary Markov chain.

math.PR↗
arXiv2019

Scale-free percolation in continuum space: quenched degree and clustering coefficient

Spatial random graphs capture several important properties of real-world networks. We prove quenched results for the continuum space version of scale-free percolation introduced in [DW18]. This is an undirected inhomogeneous random graph whose vertices are given by a Poisson point process in $\mathbb R^d$. Each vertex is equipped with a random weight and the probability that two vertices are connected by an edge depends on their weights and on their distance. Under suitable conditions on the parameters of the model, we show that, for almost all realizations of the point process, the degree distributions of all the nodes of the graph follow a power law with the same tail at infinity. We also show that the averaged clustering coefficient of the graph is self-averaging. In particular, it is almost surely equal to the the annealed clustering coefficient of one point, which is a strictly positive quantity.

math.PR↗
arXiv2018

Galton-Watson and branching process representations of the normalized Perron-Frobenius eigenvector

Let $A$ be a primitive matrix and let $λ$ be its Perron-Frobenius eigenvalue. We give formulas expressing the associated normalized Perron-Frobenius eigenvector as a simple functional of a multitype Galton-Watson process whose mean matrix is $A$, as well as of a multitype branching process with mean matrix $e^{(A-I)t}$. These formulas are generalizations of the classical formula for the invariant probability measure of a Markov chain.

math.PR↗
arXiv2017

The Wright--Fisher model for class--dependent fitness landscapes

We consider a population evolving under mutation and selection. The genotype of an individual is a word of length $\ell$ over a finite alphabet. Mutations occur during reproduction, independently on each locus; the fitness depends on the Hamming class (the distance to a reference sequence $w^*$). Evolution is driven according to the classical Wright--Fisher process. We focus on the proportion of the different classes under the invariant measure of the process. We consider the regime where the length of the genotypes $\ell$ goes to infinity, and both the population size and the inverse of the mutation rate are of order $\ell$. We prove the existence of a critical curve, which depends both on the population size and the mutation rate. Below the critical curve, the proportion of any fixed class converges to $0$, whereas above the curve, it converges to a positive quantity, for which we give an explicit formula.

math.PR↗
arXiv2017

A Markov chain representation of the Perron-Frobenius eigenvector

We consider the problem of finding the Perron-Frobenius eigenvector of a primitive matrix. Dividing each of the rows of the matrix by the sum of the elements in the row, the resulting new matrix is stochastic. We give a formula for the Perron-Frobenius eigenvector of the original matrix, in terms of a realization of the Markov chain defined by the associated stochastic matrix. This formula is a generalization of the classical formula for the invariant probability measure of a Markov chain.

math.PR↗
arXiv2017

Asymptotic behavior of Eigen's quasispecies model

We study Eigen's quasispecies model in the asymptotic regime where the length of the genotypes goes to infinity and the mutation probability goes to 0. A limiting infinite system of differential equations is obtained. We prove the convergence of the trajectories, as well as the convergence of the equilibrium solutions. We give the analogous results for a discrete-time version of Eigen's model, which coincides with a model proposed by Moran.

q-bio.PE↗
arXiv2016

The quasispecies distribution

The quasispecies model was introduced in 1971 by Manfred Eigen to discuss the first stages of life on Earth. It provides an appealing mathematical framework to study the evolution of populations in biology, for instance viruses. We present briefly the model and we focus on its stationary solutions. These formulae have a surprisingly rich combinatorial structure, involving for instance the Eulerian and Stirling numbers, as well as the up--down coefficients of permutations.

q-bio.PE↗
arXiv2016

The distribution of the quasispecies for a Galton--Watson process on the sharp peak landscape

We study a classical multitype Galton--Watson process with mutation and selection. The individuals are sequences of fixed length over a finite alphabet. On the sharp peak fitness landscape together with independent mutations per locus, we show that, as the length of the sequences goes to $\infty$ and the mutation probability goes to 0, the asymptotic relative frequency of the sequences differing on $k$ digits from the master sequence approaches $$(σe^{-a}-1)\frac{a^k}{k!}\sum_{i\geq1}\frac{i^k}{σ^i}\,,$$ where $σ$ is the selective advantage of the master sequence and $a$ is the product of the length of the chains with the mutation probability. The probability distribution $\mathcal{Q}(σ,a)$ on the non negative integers given by the above formula is the quasispecies distribution with parameters $σ$ and $a$.

math.PR↗
arXiv2016

Quasispecies on class-dependent fitness landscapes

We study Eigen's quasispecies model in the asymptotic regime where the length of the genotypes goes to infinity and the mutation probability goes to 0. We give several explicit formulas for the stationary solutions of the limiting system of differential equations.

q-bio.PE↗
arXiv2014

The distribution of the quasispecies for the Wright-Fisher model on the sharp peak landscape

We consider the classical Wright-Fisher model with mutation and selection. Mutations occur independently in each locus, and selection is performed according to the sharp peak landscape. In the asymptotic regime studied in [3], a quasispecies is formed. We find explicitly the distribution of this quasispecies, which turns out to be the same distribution as for the Moran model.

math.PR↗