Pablo Piantanida

Olivier Veilleux, Malik Boudiaf, Pablo Piantanida et al.

Transductive inference is widely used in few-shot learning, as it leverages the statistics of the unlabeled query set of a few-shot task, typically yielding substantially better performances than its inductive counterpart. The current few-shot benchmarks use perfectly class-balanced tasks at inference. We argue that such an artificial regularity is unrealistic, as it assumes that the marginal label probability of the testing samples is known and fixed to the uniform distribution. In fact, in realistic scenarios, the unlabeled query sets come with arbitrary and unknown label marginals. We introduce and study the effect of arbitrary class distributions within the query sets of few-shot tasks at inference, removing the class-balance artefact. Specifically, we model the marginal probabilities of the classes as Dirichlet-distributed random variables, which yields a principled and realistic sampling within the simplex. This leverages the current few-shot benchmarks, building testing tasks with arbitrary class distributions. We evaluate experimentally state-of-the-art transductive methods over 3 widely used data sets, and observe, surprisingly, substantial performance drops, even below inductive methods in some cases. Furthermore, we propose a generalization of the mutual-information loss, based on $α$-divergences, which can handle effectively class-distribution variations. Empirically, we show that our transductive $α$-divergence optimization outperforms state-of-the-art methods across several data sets, models and few-shot settings. Our code is publicly available at https://github.com/oveilleux/Realistic_Transductive_Few_Shot.

Ganesh Del Grosso, Hamid Jalalzai, Georg Pichler et al.

The use of personal data for training machine learning systems comes with a privacy threat and measuring the level of privacy of a model is one of the major challenges in machine learning today. Identifying training data based on a trained model is a standard way of measuring the privacy risks induced by the model. We develop a novel approach to address the problem of membership inference in pattern recognition models, relying on information provided by adversarial examples. The strategy we propose consists of measuring the magnitude of a perturbation necessary to build an adversarial example. Indeed, we argue that this quantity reflects the likelihood of belonging to the training data. Extensive numerical experiments on multivariate data and an array of state-of-the-art target models show that our method performs comparable or even outperforms state-of-the-art strategies, but without requiring any additional training samples.

Eduardo Dadalto Camara Gomes, Florence Alberge, Pierre Duhamel et al.

Reliable out-of-distribution (OOD) detection is fundamental to implementing safer modern machine learning (ML) systems. In this paper, we introduce Igeood, an effective method for detecting OOD samples. Igeood applies to any pre-trained neural network, works under various degrees of access to the ML model, does not require OOD samples or assumptions on the OOD data but can also benefit (if available) from OOD samples. By building on the geodesic (Fisher-Rao) distance between the underlying data distributions, our discriminator can combine confidence scores from the logits outputs and the learned features of a deep neural network. Empirically, we show that Igeood outperforms competing state-of-the-art methods on a variety of network architectures and datasets.

Malik Boudiaf, Etienne Bennequin, Myriam Tami et al.

We tackle the Few-Shot Open-Set Recognition (FSOSR) problem, i.e. classifying instances among a set of classes for which we only have a few labeled samples, while simultaneously detecting instances that do not belong to any known class. We explore the popular transductive setting, which leverages the unlabelled query instances at inference. Motivated by the observation that existing transductive methods perform poorly in open-set scenarios, we propose a generalization of the maximum likelihood principle, in which latent scores down-weighing the influence of potential outliers are introduced alongside the usual parametric model. Our formulation embeds supervision constraints from the support set and additional penalties discouraging overconfident predictions on the query set. We proceed with a block-coordinate descent, with the latent scores and parametric model co-optimized alternately, thereby benefiting from each other. We call our resulting formulation \textit{Open-Set Likelihood Optimization} (OSLO). OSLO is interpretable and fully modular; it can be applied on top of any pre-trained model seamlessly. Through extensive experiments, we show that our method surpasses existing inductive and transductive methods on both aspects of open-set recognition, namely inlier classification and outlier detection.

Pierre Colombo, Maxime Peyrard, Nathan Noiry et al.

Automatic evaluation metrics capable of replacing human judgments are critical to allowing fast development of new methods. Thus, numerous research efforts have focused on crafting such metrics. In this work, we take a step back and analyze recent progress by comparing the body of existing automatic metrics and human metrics altogether. As metrics are used based on how they rank systems, we compare metrics in the space of system rankings. Our extensive statistical analysis reveals surprising findings: automatic metrics -- old and new -- are much more similar to each other than to humans. Automatic metrics are not complementary and rank systems similarly. Strikingly, human metrics predict each other much better than the combination of all automatic metrics used to predict a human metric. It is surprising because human metrics are often designed to be independent, to capture different aspects of quality, e.g. content fidelity or readability. We provide a discussion of these findings and recommendations for future work in the field of evaluation.

Pierre Colombo, Victor Pellegrain, Malik Boudiaf et al.

Proprietary and closed APIs are becoming increasingly common to process natural language, and are impacting the practical applications of natural language processing, including few-shot classification. Few-shot classification involves training a model to perform a new classification task with a handful of labeled data. This paper presents three contributions. First, we introduce a scenario where the embedding of a pre-trained model is served through a gated API with compute-cost and data-privacy constraints. Second, we propose a transductive inference, a learning paradigm that has been overlooked by the NLP community. Transductive inference, unlike traditional inductive learning, leverages the statistics of unlabeled data. We also introduce a new parameter-free transductive regularizer based on the Fisher-Rao loss, which can be used on top of the gated API embeddings. This method fully utilizes unlabeled data, does not share any label with the third-party API provider and could serve as a baseline for future research. Third, we propose an improved experimental setting and compile a benchmark of eight datasets involving multiclass classification in four different languages, with up to 151 classes. We evaluate our methods using eight backbone models, along with an episodic evaluation over 1,000 episodes, which demonstrate the superiority of transductive inference over the standard inductive setting.

Pierre Colombo, Guillaume Staerman, Nathan Noiry et al.

When working with textual data, a natural application of disentangled representations is fair classification where the goal is to make predictions without being biased (or influenced) by sensitive attributes that may be present in the data (e.g., age, gender or race). Dominant approaches to disentangle a sensitive attribute from textual representations rely on learning simultaneously a penalization term that involves either an adversarial loss (e.g., a discriminator) or an information measure (e.g., mutual information). However, these methods require the training of a deep neural network with several parameter updates for each update of the representation model. As a matter of fact, the resulting nested optimization loop is both time consuming, adding complexity to the optimization dynamic, and requires a fine hyperparameter selection (e.g., learning rates, architecture). In this work, we introduce a family of regularizers for learning disentangled representations that do not require training. These regularizers are based on statistical measures of similarity between the conditional probability distributions with respect to the sensitive attributes. Our novel regularizers do not require additional training, are faster and do not involve additional tuning while achieving better results both when combined with pretrained and randomly initialized text encoders.

Maxime Darrin, Guillaume Staerman, Eduardo Dadalto Câmara Gomes et al.

Out-of-distribution (OOD) detection is a rapidly growing field due to new robustness and security requirements driven by an increased number of AI-based systems. Existing OOD textual detectors often rely on an anomaly score (e.g., Mahalanobis distance) computed on the embedding output of the last layer of the encoder. In this work, we observe that OOD detection performance varies greatly depending on the task and layer output. More importantly, we show that the usual choice (the last layer) is rarely the best one for OOD detection and that far better results could be achieved if the best layer were picked. To leverage this observation, we propose a data-driven, unsupervised method to combine layer-wise anomaly scores. In addition, we extend classical textual OOD benchmarks by including classification tasks with a greater number of classes (up to 77), which reflects more realistic settings. On this augmented benchmark, we show that the proposed post-aggregation methods achieve robust and consistent results while removing manual feature selection altogether. Their performance achieves near oracle's best layer performance.

Maxime Darrin, Pablo Piantanida, Pierre Colombo

Implementing effective control mechanisms to ensure the proper functioning and security of deployed NLP models, from translation to chatbots, is essential. A key ingredient to ensure safe system behaviour is Out-Of-Distribution (OOD) detection, which aims to detect whether an input sample is statistically far from the training distribution. Although OOD detection is a widely covered topic in classification tasks, most methods rely on hidden features output by the encoder. In this work, we focus on leveraging soft-probabilities in a black-box framework, i.e. we can access the soft-predictions but not the internal states of the model. Our contributions include: (i) RAINPROOF a Relative informAItioN Projection OOD detection framework; and (ii) a more operational evaluation setting for OOD detection. Surprisingly, we find that OOD detection is not necessarily aligned with task-specific measures. The OOD detector may filter out samples well processed by the model and keep samples that are not, leading to weaker performance. Our results show that RAINPROOF provides OOD detection methods more aligned with task-specific performance metrics than traditional OOD detectors.

Nuno M. Guerreiro, Pierre Colombo, Pablo Piantanida et al.

Neural machine translation (NMT) has become the de-facto standard in real-world machine translation applications. However, NMT models can unpredictably produce severely pathological translations, known as hallucinations, that seriously undermine user trust. It becomes thus crucial to implement effective preventive strategies to guarantee their proper functioning. In this paper, we address the problem of hallucination detection in NMT by following a simple intuition: as hallucinations are detached from the source content, they exhibit encoder-decoder attention patterns that are statistically different from those of good quality translations. We frame this problem with an optimal transport formulation and propose a fully unsupervised, plug-in detector that can be used with any attention-based NMT model. Experimental results show that our detector not only outperforms all previous model-based detectors, but is also competitive with detectors that employ large models trained on millions of samples.

Eduardo Dadalto, Marco Romanelli, Georg Pichler et al.

Misclassification detection is an important problem in machine learning, as it allows for the identification of instances where the model's predictions are unreliable. However, conventional uncertainty measures such as Shannon entropy do not provide an effective way to infer the real uncertainty associated with the model's predictions. In this paper, we introduce a novel data-driven measure of uncertainty relative to an observer for misclassification detection. By learning patterns in the distribution of soft-predictions, our uncertainty measure can identify misclassified samples based on the predicted class probabilities. Interestingly, according to the proposed measure, soft-predictions corresponding to misclassified instances can carry a large amount of uncertainty, even though they may have low Shannon entropy. We demonstrate empirical improvements over multiple image classification tasks, outperforming state-of-the-art misclassification detection methods.

Malik Boudiaf, Etienne Bennequin, Myriam Tami et al.

We tackle the Few-Shot Open-Set Recognition (FSOSR) problem, i.e. classifying instances among a set of classes for which we only have few labeled samples, while simultaneously detecting instances that do not belong to any known class. Departing from existing literature, we focus on developing model-agnostic inference methods that can be plugged into any existing model, regardless of its architecture or its training procedure. Through evaluating the embedding's quality of a variety of models, we quantify the intrinsic difficulty of model-agnostic FSOSR. Furthermore, a fair empirical evaluation suggests that the naive combination of a kNN detector and a prototypical classifier ranks before specialized or complex methods in the inductive setting of FSOSR. These observations motivated us to resort to transduction, as a popular and practical relaxation of standard few-shot learning problems. We introduce an Open Set Transductive Information Maximization method OSTIM, which hallucinates an outlier prototype while maximizing the mutual information between extracted features and assignments. Through extensive experiments spanning 5 datasets, we show that OSTIM surpasses both inductive and existing transductive methods in detecting open-set instances while competing with the strongest transductive methods in classifying closed-set instances. We further show that OSTIM's model agnosticity allows it to successfully leverage the strong expressive abilities of the latest architectures and training strategies without any hyperparameter modification, a promising sign that architectural advances to come will continue to positively impact OSTIM's performances.

Pierre Colombo, Marine Picot, Nathan Noiry et al.

The landscape of available textual adversarial attacks keeps growing, posing severe threats and raising concerns regarding the deep NLP system's integrity. However, the crucial problem of defending against malicious attacks has only drawn the attention of the NLP community. The latter is nonetheless instrumental in developing robust and trustworthy systems. This paper makes two important contributions in this line of search: (i) we introduce LAROUSSE, a new framework to detect textual adversarial attacks and (ii) we introduce STAKEOUT, a new benchmark composed of nine popular attack methods, three datasets, and two pre-trained models. LAROUSSE is ready-to-use in production as it is unsupervised, hyperparameter-free, and non-differentiable, protecting it against gradient-based methods. Our new benchmark STAKEOUT allows for a robust evaluation framework: we conduct extensive numerical experiments which demonstrate that LAROUSSE outperforms previous methods, and which allows to identify interesting factors of detection rate variations.

Pierre Colombo, Eduardo D. C. Gomes, Guillaume Staerman et al.

Deep learning methods have boosted the adoption of NLP systems in real-life applications. However, they turn out to be vulnerable to distribution shifts over time which may cause severe dysfunctions in production systems, urging practitioners to develop tools to detect out-of-distribution (OOD) samples through the lens of the neural network. In this paper, we introduce TRUSTED, a new OOD detector for classifiers based on Transformer architectures that meets operational requirements: it is unsupervised and fast to compute. The efficiency of TRUSTED relies on the fruitful idea that all hidden layers carry relevant information to detect OOD examples. Based on this, for a given input, TRUSTED consists in (i) aggregating this information and (ii) computing a similarity score by exploiting the training distribution, leveraging the powerful concept of data depth. Our extensive numerical experiments involve 51k model configurations, including various checkpoints, seeds, and datasets, and demonstrate that TRUSTED achieves state-of-the-art performances. In particular, it improves previous AUROC over 3 points.

Matthieu Dubois, François Yvon, Pablo Piantanida

The dissemination of Large Language Models (LLMs), trained at scale, and endowed with powerful text-generating abilities, has made it easier for all to produce harmful, toxic, faked or forged content. In response, various proposals have been made to automatically discriminate artificially generated from human-written texts, typically framing the problem as a binary classification problem. Early approaches evaluate an input document with a well-chosen detector LLM, assuming that low-perplexity scores reliably signal machine-made content. More recent systems instead consider two LLMs and compare their probability distributions over the document to further discriminate when perplexity alone cannot. However, using a fixed pair of models can induce brittleness in performance. We extend these approaches to the ensembling of several LLMs and derive a new, theoretically grounded approach to combine their respective strengths. Our experiments, conducted with various generator LLMs, indicate that this approach effectively leverages the strengths of each model, resulting in robust detection performance across multiple domains. Our code and data are available at https://github.com/BaggerOfWords/MOSAIC .

Federica Granese, Marine Picot, Marco Romanelli et al.

Detection of adversarial examples has been a hot topic in the last years due to its importance for safely deploying machine learning algorithms in critical applications. However, the detection methods are generally validated by assuming a single implicitly known attack strategy, which does not necessarily account for real-life threats. Indeed, this can lead to an overoptimistic assessment of the detectors' performance and may induce some bias in the comparison between competing detection schemes. We propose a novel multi-armed framework, called MEAD, for evaluating detectors based on several attack strategies to overcome this limitation. Among them, we make use of three new objectives to generate attacks. The proposed performance metric is based on the worst-case scenario: detection is successful if and only if all different attacks are correctly recognized. Empirically, we show the effectiveness of our approach. Moreover, the poor performance obtained for state-of-the-art detectors opens a new exciting line of research.

Pierre Colombo, Nathan Noiry, Guillaume Staerman et al.

One of the pursued objectives of deep learning is to provide tools that learn abstract representations of reality from the observation of multiple contextual situations. More precisely, one wishes to extract disentangled representations which are (i) low dimensional and (ii) whose components are independent and correspond to concepts capturing the essence of the objects under consideration (Locatello et al., 2019b). One step towards this ambitious project consists in learning disentangled representations with respect to a predefined (sensitive) attribute, e.g., the gender or age of the writer. Perhaps one of the main application for such disentangled representations is fair classification. Existing methods extract the last layer of a neural network trained with a loss that is composed of a cross-entropy objective and a disentanglement regularizer. In this work, we adopt an information-theoretic view of this problem which motivates a novel family of regularizers that minimizes the mutual information between the latent representation and the sensitive attribute conditional to the target. The resulting set of losses, called CLINIC, is parameter free and thus, it is easier and faster to train. CLINIC losses are studied through extensive numerical experiments by training over 2k neural networks. We demonstrate that our methods offer a better disentanglement/accuracy trade-off than previous techniques, and generalize better than training with cross-entropy loss solely provided that the disentanglement task is not too constraining.

Eduardo Dadalto, Pierre Colombo, Guillaume Staerman et al.

A key feature of out-of-distribution (OOD) detection is to exploit a trained neural network by extracting statistical patterns and relationships through the multi-layer classifier to detect shifts in the expected input data distribution. Despite achieving solid results, several state-of-the-art methods rely on the penultimate or last layer outputs only, leaving behind valuable information for OOD detection. Methods that explore the multiple layers either require a special architecture or a supervised objective to do so. This work adopts an original approach based on a functional view of the network that exploits the sample's trajectories through the various layers and their statistical dependencies. It goes beyond multivariate features aggregation and introduces a baseline rooted in functional anomaly detection. In this new framework, OOD detection translates into detecting samples whose trajectories differ from the typical behavior characterized by the training set. We validate our method and empirically demonstrate its effectiveness in OOD detection compared to strong state-of-the-art baselines on computer vision benchmarks.

Federica Granese, Marco Romanelli, Siddharth Garg et al.

Multi-armed adversarial attacks, in which multiple algorithms and objective loss functions are simultaneously used at evaluation time, have been shown to be highly successful in fooling state-of-the-art adversarial examples detectors while requiring no specific side information about the detection mechanism. By formalizing the problem at hand, we can propose a solution that aggregates the soft-probability outputs of multiple pre-trained detectors according to a minimax approach. The proposed framework is mathematically sound, easy to implement, and modular, allowing for integrating existing or future detectors. Through extensive evaluation on popular datasets (e.g., CIFAR10 and SVHN), we show that our aggregation consistently outperforms individual state-of-the-art detectors against multi-armed adversarial attacks, making it an effective solution to improve the resilience of available methods.

Mohammadhadi Shateri, Francisco Messina, Fabrice Labeau et al.

Generative Adversarial Networks (GANs) have been widely used for generating synthetic data for cases where there is a limited size real-world dataset or when data holders are unwilling to share their data samples. Recent works showed that GANs, due to overfitting and memorization, might leak information regarding their training data samples. This makes GANs vulnerable to Membership Inference Attacks (MIAs). Several defense strategies have been proposed in the literature to mitigate this privacy issue. Unfortunately, defense strategies based on differential privacy are proven to reduce extensively the quality of the synthetic data points. On the other hand, more recent frameworks such as PrivGAN and PAR-GAN are not suitable for small-size training datasets. In the present work, the overfitting in GANs is studied in terms of the discriminator, and a more general measure of overfitting based on the Bhattacharyya coefficient is defined. Then, inspired by Fano's inequality, our first defense mechanism against MIAs is proposed. This framework, which requires only a simple modification in the loss function of GANs, is referred to as the maximum entropy GAN or MEGAN and significantly improves the robustness of GANs to MIAs. As a second defense strategy, a more heuristic model based on minimizing the information leaked from generated samples about the training data points is presented. This approach is referred to as mutual information minimization GAN (MIMGAN) and uses a variational representation of the mutual information to minimize the information that a synthetic sample might leak about the whole training data set. Applying the proposed frameworks to some commonly used data sets against state-of-the-art MIAs reveals that the proposed methods can reduce the accuracy of the adversaries to the level of random guessing accuracy with a small reduction in the quality of the synthetic data samples.

Lautaro Estienne, Erik Ernst, Matías Vera et al.

In high-stakes automated decision-making, access to predictive uncertainty is essential for enabling users -- human or downstream systems -- to accept or reject predictions based on application-specific cost trade-offs. Such uncertainty-augmented (UA) systems -- i.e., systems that output both predictions and uncertainty scores -- are currently being assessed in the literature in a variety of ways, using separate metrics to evaluate the predictions and the uncertainty scores, setting a cost function with a fixed rejection cost or integrating over a coverage-risk curve. We argue that these evaluation approaches are inadequate for assessing overall performance of the UA system for decision making under uncertainty and propose a novel family of metrics, \ECUAS{n}, formulated as proper scoring rules for the task of interest. The parameter $n$ controls the trade-off between the cost of incorrect predictions and imperfect uncertainties depending on the needs of the use-case. We demonstrate the advantages of the \ECUAS{n} metrics both theoretically and empirically, through experiments on diverse classification and generation datasets, including a manually annotated subset of TriviaQA.

Pablo Morales-Álvarez, Stergios Christodoulidis, Maria Vakalopoulou et al.

The emergence of large pre-trained vision-language models (VLMs) represents a paradigm shift in machine learning, with unprecedented results in a broad span of visual recognition tasks. CLIP, one of the most popular VLMs, has exhibited remarkable zero-shot and transfer learning capabilities in classification. To transfer CLIP to downstream tasks, adapters constitute a parameter-efficient approach that avoids backpropagation through the large model (unlike related prompt learning methods). However, CLIP adapters have been developed to target discriminative performance, and the quality of their uncertainty estimates has been overlooked. In this work we show that the discriminative performance of state-of-the-art CLIP adapters does not always correlate with their uncertainty estimation capabilities, which are essential for a safe deployment in real-world scenarios. We also demonstrate that one of such adapters is obtained through MAP inference from a more general probabilistic framework. Based on this observation we introduce BayesAdapter, which leverages Bayesian inference to estimate a full probability distribution instead of a single point, better capturing the variability inherent in the parameter space. In a comprehensive empirical evaluation we show that our approach obtains high quality uncertainty estimates in the predictions, standing out in calibration and selective classification. Our code will be publicly available upon acceptance of the paper.

Matthieu Dubois, François Yvon, Pablo Piantanida

As texts generated by Large Language Models (LLMs) are ever more common and often indistinguishable from human-written content, research on automatic text detection has attracted growing attention. Many recent detectors report near-perfect accuracy, often boasting AUROC scores above 99\%. However, these claims typically assume fixed generation settings, leaving open the question of how robust such systems are to changes in decoding strategies. In this work, we systematically examine how sampling-based decoding impacts detectability, with a focus on how subtle variations in a model's (sub)word-level distribution affect detection performance. We find that even minor adjustments to decoding parameters - such as temperature, top-p, or nucleus sampling - can severely impair detector accuracy, with AUROC dropping from near-perfect levels to 1\% in some settings. Our findings expose critical blind spots in current detection methods and emphasize the need for more comprehensive evaluation protocols. To facilitate future research, we release a large-scale dataset encompassing 37 decoding configurations, along with our code and evaluation framework https://github.com/BaggerOfWords/Sampling-and-Detection

Pierre Marza, Leo Fillioux, Sofiène Boutaj et al.

Progress in a research field can be hard to assess, in particular when many concurrent methods are proposed in a short period of time. This is the case in digital pathology, where many foundation models have been released recently to serve as feature extractors for tile-level images, being used in a variety of downstream tasks, both for tile- and slide-level problems. Benchmarking available methods then becomes paramount to get a clearer view of the research landscape. In particular, in critical domains such as healthcare, a benchmark should not only focus on evaluating downstream performance, but also provide insights about the main differences between methods, and importantly, further consider uncertainty and robustness to ensure a reliable usage of proposed models. For these reasons, we introduce THUNDER, a tile-level benchmark for digital pathology foundation models, allowing for efficient comparison of many models on diverse datasets with a series of downstream tasks, studying their feature spaces and assessing the robustness and uncertainty of predictions informed by their embeddings. THUNDER is a fast, easy-to-use, dynamic benchmark that can already support a large variety of state-of-the-art foundation, as well as local user-defined models for direct tile-based comparison. In this paper, we provide a comprehensive comparison of 23 foundation models on 16 different datasets covering diverse tasks, feature analysis, and robustness. The code for THUNDER is publicly available at https://github.com/MICS-Lab/thunder.

Georg Pichler, Marco Romanelli, Pablo Piantanida

In this paper, we propose a novel fairness framework grounded in the concept of happiness, a measure of the utility each group gains fromdecisionoutcomes. Bycapturingfairness through this intuitive lens, we not only offer a more human-centered approach, but also one that is mathematically rigorous: In order to compute the optimal, fair post-processing strategy, only a linear program needs to be solved. This makes our method both efficient and scalable with existing optimization tools. Furthermore, it unifies and extends several well-known fairness definitions, and our empirical results highlight its practical strengths across diverse scenarios.

Eric Aubinais, Elisabeth Gassiat, Pablo Piantanida

Membership inference attacks (MIA) can reveal whether a particular data point was part of the training dataset, potentially exposing sensitive information about individuals. This article provides theoretical guarantees by exploring the fundamental statistical limitations associated with MIAs on machine learning models at large. More precisely, we first derive the statistical quantity that governs the effectiveness and success of such attacks. We then theoretically prove that in a non-linear regression setting with overfitting learning procedures, attacks may have a high probability of success. Finally, we investigate several situations for which we provide bounds on this quantity of interest. Interestingly, our findings indicate that discretizing the data might enhance the learning procedure's security. Specifically, it is demonstrated to be limited by a constant, which quantifies the diversity of the underlying data distribution. We illustrate those results through simple simulations.

Philippe Formont, Maxime Darrin, Ismail Ben Ayed et al.

Recent advances in reasoning-based large language models (LLMs) have demonstrated substantial improvements in complex problem-solving tasks. Motivated by these advances, several works have explored the application of reasoning LLMs to drug discovery and molecular design. However, most existing approaches either focus on evaluation or rely on training setups that require ground-truth labels, such as molecule pairs with known property modifications. Such supervision is unavailable in \textit{de novo} molecular generation, where the objective is to generate novel molecules that optimize a desirability score without prior knowledge of high-scoring candidates. To bridge this gap, we introduce MolRGen, a large-scale benchmark and dataset for training and evaluating reasoning-based LLMs on \textit{de novo} molecular generation. Our contributions are threefold. First, we propose a setting to evaluate and train models for \textit{de novo} molecular generation and property prediction. Second, we introduce a novel diversity-aware top-$k$ score that captures both the quality and diversity of generated molecules. Third, we show our setting can be used to train LLMs for molecular generation, training a 24B LLM with reinforcement learning, and we provide a detailed analysis of its performance and limitations.

Arthur Satouf, Gabriel Ben Zenou, Benjamin Piwowarski et al.

Current sparse neural information retrieval (IR) methods, and to a lesser extent more traditional models such as BM25, do not take into account the document collection and the complex interplay between different term weights when representing a single document. In this paper, we show how the Rational Speech Acts (RSA), a linguistics framework used to minimize the number of features to be communicated when identifying an object in a set, can be adapted to the IR case -- and in particular to the high number of potential features (here, tokens). RSA dynamically modulates token-document interactions by considering the influence of other documents in the dataset, better contrasting document representations. Experiments show that incorporating RSA consistently improves multiple sparse retrieval models and achieves state-of-the-art performance on out-of-domain datasets from the BEIR benchmark. https://github.com/arthur-75/Rational-Retrieval-Acts

Eduardo Dadalto, Florence Alberge, Pierre Duhamel et al.

This paper introduces a universal approach to seamlessly combine out-of-distribution (OOD) detection scores. These scores encompass a wide range of techniques that leverage the self-confidence of deep learning models and the anomalous behavior of features in the latent space. Not surprisingly, combining such a varied population using simple statistics proves inadequate. To overcome this challenge, we propose a quantile normalization to map these scores into p-values, effectively framing the problem into a multi-variate hypothesis test. Then, we combine these tests using established meta-analysis tools, resulting in a more effective detector with consolidated decision boundaries. Furthermore, we create a probabilistic interpretable criterion by mapping the final statistics into a distribution with known parameters. Through empirical investigation, we explore different types of shifts, each exerting varying degrees of impact on data. Our results demonstrate that our approach significantly improves overall robustness and performance across diverse OOD detection scenarios. Notably, our framework is easily extensible for future developments in detection scores and stands as the first to combine decision boundaries in this context. The code and artifacts associated with this work are publicly available\footnote{\url{https://github.com/edadaltocg/detectors}}.

Andres Altieri, Marco Romanelli, Georg Pichler et al.

This paper tackles the challenge of detecting unreliable behavior in regression algorithms, which may arise from intrinsic variability (e.g., aleatoric uncertainty) or modeling errors (e.g., model uncertainty). First, we formally introduce the notion of unreliability in regression, i.e., when the output of the regressor exceeds a specified discrepancy (or error). Then, using powerful tools for probabilistic modeling, we estimate the discrepancy density, and we measure its statistical diversity using our proposed metric for statistical dissimilarity. In turn, this allows us to derive a data-driven score that expresses the uncertainty of the regression outcome. We show empirical improvements in error detection for multiple regression tasks, consistently outperforming popular baseline approaches, and contributing to the broader field of uncertainty quantification and safe machine learning systems. Our code is available at https://zenodo.org/records/11281964.

Roberta Hansen, Matias Vera, Lautaro Estienne et al.

This paper deals with the convergence analysis of the SUCPA (Semi Unsupervised Calibration through Prior Adaptation) algorithm, defined from a first-order non-linear difference equations, first developed to correct the scores output by a supervised machine learning classifier. The convergence analysis is addressed as a dynamical system problem, by studying the local and global stability of the nonlinear map derived from the algorithm. This map, which is defined by a composition of exponential and rational functions, turns out to be non-hyperbolic with a non-bounded set of non-isolated fixed points. Hence, a non-standard method for solving the convergence analysis is used consisting of an ad-hoc geometrical approach. For a binary classification problem (two-dimensional map), we rigorously prove that the map is globally asymptotically stable. Numerical experiments on real-world application are performed to support the theoretical results by means of two different classification problems: Sentiment Polarity performed with a Large Language Model and Cat-Dog Image classification. For a greater number of classes, the numerical evidence shows the same behavior of the algorithm, and this is illustrated with a Natural Language Inference example. The experiment codes are publicly accessible online at the following repository: https://github.com/LautaroEst/sucpa-convergence

Georg Pichler, Pierre Colombo, Malik Boudiaf et al.

Mutual Information (MI) has been widely used as a loss regularizer for training neural networks. This has been particularly effective when learn disentangled or compressed representations of high dimensional data. However, differential entropy (DE), another fundamental measure of information, has not found widespread use in neural network training. Although DE offers a potentially wider range of applications than MI, off-the-shelf DE estimators are either non differentiable, computationally intractable or fail to adapt to changes in the underlying distribution. These drawbacks prevent them from being used as regularizers in neural networks training. To address shortcomings in previously proposed estimators for DE, here we introduce KNIFE, a fully parameterized, differentiable kernel-based estimator of DE. The flexibility of our approach also allows us to construct KNIFE-based estimators for conditional (on either discrete or continuous variables) DE, as well as MI. We empirically validate our method on high-dimensional synthetic data and further apply it to guide the training of neural networks for real-world tasks. Our experiments on a large variety of tasks, including visual domain adaptation, textual fair classification, and textual fine-tuning demonstrate the effectiveness of KNIFE-based estimation. Code can be found at https://github.com/g-pichler/knife.

Malik Boudiaf, Ziko Imtiaz Masud, Jérôme Rony et al.

We introduce Transductive Infomation Maximization (TIM) for few-shot learning. Our method maximizes the mutual information between the query features and their label predictions for a given few-shot task, in conjunction with a supervision loss based on the support set. We motivate our transductive loss by deriving a formal relation between the classification accuracy and mutual-information maximization. Furthermore, we propose a new alternating-direction solver, which substantially speeds up transductive inference over gradient-based optimization, while yielding competitive accuracy. We also provide a convergence analysis of our solver based on Zangwill's theory and bound-optimization arguments. TIM inference is modular: it can be used on top of any base-training feature extractor. Following standard transductive few-shot settings, our comprehensive experiments demonstrate that TIM outperforms state-of-the-art methods significantly across various datasets and networks, while used on top of a fixed feature extractor trained with simple cross-entropy on the base classes, without resorting to complex meta-learning schemes. It consistently brings between 2 % and 5 % improvement in accuracy over the best performing method, not only on all the well-established few-shot benchmarks but also on more challenging scenarios, with random tasks, domain shift and larger numbers of classes, as in the recently introduced META-DATASET. Our code is publicly available at https://github.com/mboudiaf/TIM. We also publicly release a standalone PyTorch implementation of META-DATASET, along with additional benchmarking results, at https://github.com/mboudiaf/pytorch-meta-dataset.

Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud et al.

We show that the way inference is performed in few-shot segmentation tasks has a substantial effect on performances -- an aspect often overlooked in the literature in favor of the meta-learning paradigm. We introduce a transductive inference for a given query image, leveraging the statistics of its unlabeled pixels, by optimizing a new loss containing three complementary terms: i) the cross-entropy on the labeled support pixels; ii) the Shannon entropy of the posteriors on the unlabeled query-image pixels; and iii) a global KL-divergence regularizer based on the proportion of the predicted foreground. As our inference uses a simple linear classifier of the extracted features, its computational load is comparable to inductive inference and can be used on top of any base training. Foregoing episodic training and using only standard cross-entropy training on the base classes, our inference yields competitive performances on standard benchmarks in the 1-shot scenarios. As the number of available shots increases, the gap in performances widens: on PASCAL-5i, our method brings about 5% and 6% improvements over the state-of-the-art, in the 5- and 10-shot scenarios, respectively. Furthermore, we introduce a new setting that includes domain shifts, where the base and novel classes are drawn from different datasets. Our method achieves the best performances in this more realistic setting. Our code is freely available online: https://github.com/mboudiaf/RePRI-for-Few-Shot-Segmentation.

Arthur Satouf, Yuxuan Zong, Habiboulaye Amadou-Boubacar et al.

Generative Retrieval (GR) differs from the traditional index-then-retrieve pipeline by storing relevance in model parameters and directly generating document identifiers. However, GR often struggles to generalize and is costly to scale. We introduce QUESTER (QUEry SpecificaTion gEnerative Retrieval), which reframes GR as query specification generation - in this work, a simple keyword query handled by BM25 - using a (small) LLM. The policy is trained using reinforcement learning techniques (GRPO). Across in- and out-of-domain evaluations, we show that our model is more effective than BM25, and competitive with neural IR models, while maintaining a good efficiency

Lautaro Estienne, Luciana Ferrer, Matías Vera et al.

A wide variety of natural language tasks are currently being addressed with large-scale language models (LLMs). These models are usually trained with a very large amount of unsupervised text data and adapted to perform a downstream natural language task using methods like fine-tuning, calibration or in-context learning. In this work, we propose an approach to adapt the prior class distribution to perform text classification tasks without the need for labelled samples and only few in-domain sample queries. The proposed approach treats the LLM as a black box, adding a stage where the model posteriors are calibrated to the task. Results show that these methods outperform the un-adapted model for different number of training shots in the prompt and a previous approach were calibration is performed without using any adaptation data.

Maxime Darrin, Philippe Formont, Jackie Chi Kit Cheung et al.

Assessing the quality of summarizers poses significant challenges. In response, we propose a novel task-oriented evaluation approach that assesses summarizers based on their capacity to produce summaries that are useful for downstream tasks, while preserving task outcomes. We theoretically establish a direct relationship between the resulting error probability of these tasks and the mutual information between source texts and generated summaries. We introduce $\texttt{COSMIC}$ as a practical implementation of this metric, demonstrating its strong correlation with human judgment-based metrics and its effectiveness in predicting downstream task performance. Comparative analyses against established metrics like $\texttt{BERTScore}$ and $\texttt{ROUGE}$ highlight the competitive performance of $\texttt{COSMIC}$.

Loïc Fosse, Frédéric Béchet, Benoît Favre et al.

Tasks are central in machine learning, as they are the most natural objects to assess the capabilities of current models. The trend is to build general models able to address any task. Even though transfer learning and multitask learning try to leverage the underlying task space, no well-founded tools are available to study its structure. This study proposes a theoretically grounded setup to define the notion of task and to compute the {\bf inclusion} between two tasks from a statistical deficiency point of view. We propose a tractable proxy as information sufficiency to estimate the degree of inclusion between tasks, show its soundness on synthetic data, and use it to reconstruct empirically the classic NLP pipeline.

Eric Aubinais, Philippe Formont, Pablo Piantanida et al.

Quantizing machine learning models has demonstrated its effectiveness in lowering memory and inference costs while maintaining performance levels comparable to the original models. In this work, we investigate the impact of quantization procedures on the privacy of data-driven models, specifically focusing on their vulnerability to membership inference attacks. We derive an asymptotic theoretical analysis of Membership Inference Security (MIS), characterizing the privacy implications of quantized algorithm weights against the most powerful (and possibly unknown) attacks. Building on these theoretical insights, we propose a novel methodology to empirically assess and rank the privacy levels of various quantization procedures. Using synthetic datasets, we demonstrate the effectiveness of our approach in assessing the MIS of different quantizers. Furthermore, we explore the trade-off between privacy and performance using real-world data and models in the context of molecular modeling.

Philippe Formont, Hugo Jeannin, Pablo Piantanida et al.

Few-shot learning has recently attracted significant interest in drug discovery, with a recent, fast-growing literature mostly involving convoluted meta-learning strategies. We revisit the more straightforward fine-tuning approach for molecular data, and propose a regularized quadratic-probe loss based on the the Mahalanobis distance. We design a dedicated block-coordinate descent optimizer, which avoid the degenerate solutions of our loss. Interestingly, our simple fine-tuning approach achieves highly competitive performances in comparison to state-of-the-art methods, while being applicable to black-box settings and removing the need for specific episodic pre-training strategies. Furthermore, we introduce a new benchmark to assess the robustness of the competing methods to domain shifts. In this setting, our fine-tuning baseline obtains consistently better results than meta-learning methods.

Philippe Formont, Maxime Darrin, Banafsheh Karimian et al.

Casting complex inputs into tractable representations is a critical step across various fields. Diverse embedding models emerge from differences in architectures, loss functions, input modalities and datasets, each capturing unique aspects of the input. Multi-teacher distillation leverages this diversity to enrich representations but often remains tailored to specific tasks. In this paper, we introduce a task-agnostic framework based on a ``majority vote" objective function. We demonstrate that this function is bounded by the mutual information between student and teachers' embeddings, leading to a task-agnostic distillation loss that eliminates dependence on task-specific labels or prior knowledge. Our evaluations across text, vision models, and molecular modeling show that our method effectively leverages teacher diversity, resulting in representations enabling better performance for a wide range of downstream tasks such as classification, clustering, or regression. Additionally, we train and release state-of-the-art embedding models, enhancing downstream performance in various modalities.

Lautaro Estienne, Gabriel Ben Zenou, Nona Naderi et al.

As AI systems take on collaborative roles, they must reason about shared goals and beliefs-not just generate fluent language. The Rational Speech Act (RSA) framework offers a principled approach to pragmatic reasoning, but existing extensions face challenges in scaling to multi-turn, collaborative scenarios. In this paper, we introduce Collaborative Rational Speech Act (CRSA), an information-theoretic (IT) extension of RSA that models multi-turn dialog by optimizing a gain function adapted from rate-distortion theory. This gain is an extension of the gain model that is maximized in the original RSA model but takes into account the scenario in which both agents in a conversation have private information and produce utterances conditioned on the dialog. We demonstrate the effectiveness of CRSA on referential games and template-based doctor-patient dialogs in the medical domain. Empirical results show that CRSA yields more consistent, interpretable, and collaborative behavior than existing baselines-paving the way for more pragmatic and socially aware language agents.

Cesare Spinoso-Di Piano, David Austin, Pablo Piantanida et al.

Figurative language (e.g., irony, hyperbole, understatement) is ubiquitous in human communication, resulting in utterances where the literal and the intended meanings do not match. The Rational Speech Act (RSA) framework, which explicitly models speaker intentions, is the most widespread theory of probabilistic pragmatics, but existing implementations are either unable to account for figurative expressions or require modeling the implicit motivations for using figurative language (e.g., to express joy or annoyance) in a setting-specific way. In this paper, we introduce the Rhetorical-Strategy-Aware RSA $(RSA)^2$ framework which models figurative language use by considering a speaker's employed rhetorical strategy. We show that $(RSA)^2$ enables human-compatible interpretations of non-literal utterances without modeling a speaker's motivations for being non-literal. Combined with LLMs, it achieves state-of-the-art performance on the ironic split of PragMega+, a new irony interpretation dataset introduced in this study.

Florence Regol, Joud Chataoui, Bertrand Charpentier et al.

Machine learning models can solve complex tasks but often require significant computational resources during inference. This has led to the development of various post-training computation reduction methods that tackle this issue in different ways, such as quantization which reduces the precision of weights and arithmetic operations, and dynamic networks which adapt computation to the sample at hand. In this work, we propose a more general dynamic network that can combine both quantization and early exit dynamic network: QuEE. Our algorithm can be seen as a form of soft early exiting or input-dependent compression. Rather than a binary decision between exiting or continuing, we introduce the possibility of continuing with reduced computation. This complicates the traditionally considered early exiting problem, which we solve through a principled formulation. The crucial factor of our approach is accurate prediction of the potential accuracy improvement achievable through further computation. We demonstrate the effectiveness of our method through empirical evaluation, as well as exploring the conditions for its success on 4 classification datasets.

Maxime Darrin, Philippe Formont, Ismail Ben Ayed et al.

Embedders play a central role in machine learning, projecting any object into numerical representations that can, in turn, be leveraged to perform various downstream tasks. The evaluation of embedding models typically depends on domain-specific empirical approaches utilizing downstream tasks, primarily because of the lack of a standardized framework for comparison. However, acquiring adequately large and representative datasets for conducting these assessments is not always viable and can prove to be prohibitively expensive and time-consuming. In this paper, we present a unified approach to evaluate embedders. First, we establish theoretical foundations for comparing embedding models, drawing upon the concepts of sufficiency and informativeness. We then leverage these concepts to devise a tractable comparison criterion (information sufficiency), leading to a task-agnostic and self-supervised ranking procedure. We demonstrate experimentally that our approach aligns closely with the capability of embedding models to facilitate various downstream tasks in both natural language processing and molecular biology. This effectively offers practitioners a valuable tool for prioritizing model trials.

Maxime Darrin, Ines Arous, Pablo Piantanida et al.

Scientific peer review is essential for the quality of academic publications. However, the increasing number of paper submissions to conferences has strained the reviewing process. This surge poses a burden on area chairs who have to carefully read an ever-growing volume of reviews and discern each reviewer's main arguments as part of their decision process. In this paper, we introduce \sys, a summarization method designed to offer a concise yet comprehensive overview of scholarly reviews. Unlike traditional consensus-based methods, \sys extracts both common and unique opinions from the reviews. We introduce novel uniqueness scores based on the Rational Speech Act framework to identify relevant sentences in the reviews. Our method aims to provide a pragmatic glimpse into all reviews, offering a balanced perspective on their opinions. Our experimental results with both automatic metrics and human evaluation show that \sys generates more discriminative summaries than baseline methods in terms of human evaluation while achieving comparable performance with these methods in terms of automatic metrics.

Federica Granese, Marco Romanelli, Pablo Piantanida

This paper explores a scenario in which a malicious actor employs a multi-armed attack strategy to manipulate data samples, offering them various avenues to introduce noise into the dataset. Our central objective is to protect the data by detecting any alterations to the input. We approach this defensive strategy with utmost caution, operating in an environment where the defender possesses significantly less information compared to the attacker. Specifically, the defender is unable to utilize any data samples for training a defense model or verifying the integrity of the channel. Instead, the defender relies exclusively on a set of pre-existing detectors readily available "off the shelf". To tackle this challenge, we derive an innovative information-theoretic defense approach that optimally aggregates the decisions made by these detectors, eliminating the need for any training data. We further explore a practical use-case scenario for empirical evaluation, where the attacker possesses a pre-trained classifier and launches well-known adversarial attacks against it. Our experiments highlight the effectiveness of our proposed solution, even in scenarios that deviate from the optimal setup.

Georg Pichler, Marco Romanelli, Leonardo Rey Vega et al.

Federated Learning is expected to provide strong privacy guarantees, as only gradients or model parameters but no plain text training data is ever exchanged either between the clients or between the clients and the central server. In this paper, we challenge this claim by introducing a simple but still very effective membership inference attack algorithm, which relies only on a single training step. In contrast to the popular honest-but-curious model, we investigate a framework with a dishonest central server. Our strategy is applicable to models with ReLU activations and uses the properties of this activation function to achieve perfect accuracy. Empirical evaluation on visual classification tasks with MNIST, CIFAR10, CIFAR100 and CelebA datasets show that our method provides perfect accuracy in identifying one sample in a training set with thousands of samples. Occasional failures of our method lead us to discover duplicate images in the CIFAR100 and CelebA datasets.

Matias Vera, Leonardo Rey Vega, Pablo Piantanida

The ultimate performance of machine learning algorithms for classification tasks is usually measured in terms of the empirical error probability (or accuracy) based on a testing dataset. Whereas, these algorithms are optimized through the minimization of a typically different--more convenient--loss function based on a training set. For classification tasks, this loss function is often the negative log-loss that leads to the well-known cross-entropy risk which is typically better behaved (from a numerical perspective) than the error probability. Conventional studies on the generalization error do not usually take into account the underlying mismatch between losses at training and testing phases. In this work, we introduce an analysis based on point-wise PAC approach over the generalization gap considering the mismatch of testing based on the accuracy metric and training on the negative log-loss. We label this analysis PACMAN. Building on the fact that the mentioned mismatch can be written as a likelihood ratio, concentration inequalities can be used to provide some insights for the generalization problem in terms of some point-wise PAC bounds depending on some meaningful information-theoretic quantities. An analysis of the obtained bounds and a comparison with available results in the literature are also provided.

Pierre Colombo, Chloe Clavel, Pablo Piantanida

Assessing the quality of natural language generation systems through human annotation is very expensive. Additionally, human annotation campaigns are time-consuming and include non-reusable human labour. In practice, researchers rely on automatic metrics as a proxy of quality. In the last decade, many string-based metrics (e.g., BLEU) have been introduced. However, such metrics usually rely on exact matches and thus, do not robustly handle synonyms. In this paper, we introduce InfoLM a family of untrained metrics that can be viewed as a string-based metric that addresses the aforementioned flaws thanks to a pre-trained masked language model. This family of metrics also makes use of information measures allowing the adaptation of InfoLM to various evaluation criteria. Using direct assessment, we demonstrate that InfoLM achieves statistically significant improvement and over $10$ points of correlation gains in many configurations on both summarization and data2text generation.