arXiv subjects

Bram Wouters

Publications and source records attributed to Bram Wouters.

11 recordsLinked to original sources

arXiv2025

Optimizing importance weighting in the presence of sub-population shifts

A distribution shift between the training and test data can severely harm performance of machine learning models. Importance weighting addresses this issue by assigning different weights to data points during training. We argue that existing heuristics for determining the weights are suboptimal, as they neglect the increase of the variance of the estimated model due to the finite sample size of the training data. We interpret the optimal weights in terms of a bias-variance trade-off, and propose a bi-level optimization procedure in which the weights and model parameters are optimized simultaneously. We apply this optimization to existing importance weighting techniques for last-layer retraining of deep neural networks in the presence of sub-population shifts and show empirically that optimizing weights significantly improves generalization performance.

stat.ML↗

arXiv2025

Preserving Task-Relevant Information Under Linear Concept Removal

Modern neural networks often encode unwanted concepts alongside task-relevant information, leading to fairness and interpretability concerns. Existing post-hoc approaches can remove undesired concepts but often degrade useful signals. We introduce SPLINCE-Simultaneous Projection for LINear concept removal and Covariance prEservation - which eliminates sensitive concepts from representations while exactly preserving their covariance with a target label. SPLINCE achieves this via an oblique projection that 'splices out' the unwanted direction yet protects important label correlations. Theoretically, it is the unique solution that removes linear concept predictability and maintains target covariance with minimal embedding distortion. Empirically, SPLINCE outperforms baselines on benchmarks such as Bias in Bios and Winobias, removing protected attributes while minimally damaging main-task information.

cs.LG↗

arXiv2024

Removing Spurious Concepts from Neural Network Representations via Joint Subspace Estimation

Out-of-distribution generalization in neural networks is often hampered by spurious correlations. A common strategy is to mitigate this by removing spurious concepts from the neural network representation of the data. Existing concept-removal methods tend to be overzealous by inadvertently eliminating features associated with the main task of the model, thereby harming model performance. We propose an iterative algorithm that separates spurious from main-task concepts by jointly identifying two low-dimensional orthogonal subspaces in the neural network representation. We evaluate the algorithm on benchmark datasets for computer vision (Waterbirds, CelebA) and natural language processing (MultiNLI), and show that it outperforms existing concept removal methods

cs.LG↗

arXiv2023

Optimizing watermarks for large language models

With the rise of large language models (LLMs) and concerns about potential misuse, watermarks for generative LLMs have recently attracted much attention. An important aspect of such watermarks is the trade-off between their identifiability and their impact on the quality of the generated text. This paper introduces a systematic approach to this trade-off in terms of a multi-objective optimization problem. For a large class of robust, efficient watermarks, the associated Pareto optimal solutions are identified and shown to outperform the currently default watermark.

cs.CR↗

arXiv2023

Noise reduction for functional time series

A novel method for noise reduction in the setting of curve time series with error contamination is proposed, based on extending the framework of functional principal component analysis (FPCA). We employ the underlying, finite-dimensional dynamics of the functional time series to separate the serially dependent dynamical part of the observed curves from the noise. Upon identifying the subspaces of the signal and idiosyncratic components, we construct a projection of the observed curve time series along the noise subspace, resulting in an estimate of the underlying denoised curves. This projection is optimal in the sense that it minimizes the mean integrated squared error. By applying our method to similated and real data, we show the denoising estimator is consistent and outperforms existing denoising techniques. Furthermore, we show it can be used as a pre-processing step to improve forecasting.

stat.ME↗

arXiv2015

Complete Generalized Gibbs Ensemble in an interacting Theory

In integrable many-particle systems, it is widely believed that the stationary state reached at late times after a quantum quench can be described by a generalized Gibbs ensemble (GGE) constructed from their extensive number of conserved charges. A crucial issue is then to identify a complete set of these charges, enabling the GGE to provide exact steady state predictions. Here we solve this long-standing problem for the case of the spin-1/2 Heisenberg chain by explicitly constructing a GGE which uniquely fixes the macrostate describing the stationary behaviour after a general quantum quench. A crucial ingredient in our method, which readily generalizes to other integrable models, are recently discovered quasi-local charges. As a test, we reproduce the exact post-quench steady state of the Neel quench problem obtained previously by means of the Quench Action method.

quant-ph↗

arXiv2014

Quench action approach for releasing the Néel state into the spin-1/2 XXZ chain

The steady state after a quantum quench from the Néel state to the anisotropic Heisenberg model for spin chains is investigated. Two methods that aim to describe the postquench non-thermal equilibrium, the generalized Gibbs ensemble and the quench action approach, are discussed and contrasted. Using the recent implementation of the quench action approach for this Néel-to-XXZ quench, we obtain an exact description of the steady state in terms of Bethe root densities, for which we give explicit analytical expressions. Furthermore, by developing a systematic small-quench expansion around the antiferromagnetic Ising limit, we analytically investigate the differences between the predictions of the two methods in terms of densities and postquench equilibrium expectation values of local physical observables. Finally, we discuss the details of the quench action solution for the quench to the isotropic Heisenberg spin chain. For this case we validate the underlying assumptions of the quench action approach by studying the large-system-size behavior of the overlaps between Bethe states and the Néel state.

cond-mat.str-el↗

arXiv2014

Quenching the Anisotropic Heisenberg Chain: Exact Solution and Generalized Gibbs Ensemble Predictions

We study quenches in integrable spin-1/2 chains in which we evolve the ground state of the antiferromagnetic Ising model with the anisotropic Heisenberg Hamiltonian. For this nontrivially interacting situation, an application of the first-principles-based quench action method allows us to give an exact description of the postquench steady state in the thermodynamic limit. We show that a generalized Gibbs ensemble, implemented using all known local conserved charges, fails to reproduce the exact quench action steady state and to correctly predict postquench equilibrium expectation values of physical observables. This is supported by numerical linked-cluster calculations within the diagonal ensemble in the thermodynamic limit.

cond-mat.str-el↗

arXiv2014

Néel-XXZ state overlaps: odd particle numbers and Lieb-Liniger scaling limit

We specialize a recently-proposed determinant formula for the overlap of the zero-momentum Néel state with Bethe states of the spin-1/2 XXZ chain to the case of an odd number of downturned spins, showing that it is still of "Gaudin-like" form, similar to the case of an even number of down spins. We generalize this result to the overlap of $q$-raised Néel states with parity-invariant Bethe states lying in a nonzero magnetization sector. The generalized determinant expression can then be used to derive the corresponding determinants and their prefactors in the scaling limit to the Lieb-Liniger (LL) Bose gas. The odd number of down spins directly translates to an odd number of bosons. We furthermore give a proof that the Néel state has no overlap with non-parity-invariant Bethe states. This is based on a determinant expression for overlaps with general Bethe states that was obtained in the context of the XXZ chain with open boundary conditions. The statement that overlaps with non-parity-invariant Bethe states vanish is still valid in the scaling limit to LL which means that the BEC state has zero overlap with non-parity-invariant LL Bethe states.

cond-mat.stat-mech↗

arXiv2014

A Gaudin-like determinant for overlaps of Néel and XXZ Bethe states

We derive a determinant expression for overlaps of Bethe states of the XXZ spin chain with the N{é}el state, the ground state of the system in the antiferromagnetic Ising limit. Our formula, of determinant form, is valid for generic system size. Interestingly, it is remarkably similar to the well-known Gaudin formula for the norm of Bethe states, and to another recently-derived overlap formula appearing in the Lieb-Liniger model.

cond-mat.stat-mech↗

arXiv2014

Solution for an interaction quench in the Lieb-Liniger Bose gas

We study a quench protocol where the ground state of a free many-particle bosonic theory in one dimension is let unitarily evolve in time under the integrable Lieb-Liniger Hamiltonian of $δ$-interacting repulsive bosons. By using a recently-proposed variational method, we here obtain the exact non-thermal steady-state of the system in the thermodynamic limit, and discuss some of its main physical properties. Besides being a rare case of a thermodynamically exact solution to a truly interacting quench situation, this interestingly represents an example where a naive implementation of the generalized Gibbs ensemble fails.

cond-mat.stat-mech↗