A two-head loss function for deep Average-K classification

Average-K classification is an alternative to top-K classification in which the number of labels returned varies with the ambiguity of the input image but must average to K over all the samples. A simple method to solve this task is to threshold the softmax output of a model trained with the cross-entropy loss. This approach is theoretically proven to be asymptotically consistent, but it is not guaranteed to be optimal for a finite set of samples. In this paper, we propose a new loss function based on a multi-label classification head in addition to the classical softmax. This second head is trained using pseudo-labels generated by thresholding the softmax head while guaranteeing that K classes are returned on average. We show that this approach allows the model to better capture ambiguities between classes and, as a result, to return more consistent sets of possible classes. Experiments on two datasets from the literature demonstrate that our approach outperforms the softmax baseline, as well as several other loss functions more generally designed for weakly supervised multi-label classification. The gains are larger the higher the uncertainty, especially for classes with few samples.

Improve learning combining crowdsourced labels by weighting Areas Under the Margin

In supervised learning -- for instance in image classification -- modern massive datasets are commonly labeled by a crowd of workers. The obtained labels in this crowdsourcing setting are then aggregated for training. The aggregation step generally leverages a per worker trust score. Yet, such worker-centric approaches discard each task ambiguity. Some intrinsically ambiguous tasks might even fool expert workers, which could eventually be harmful for the learning step. In a standard supervised learning setting -- with one label per task and balanced classes -- the Area Under the Margin (AUM) statistic is tailored to identify mislabeled data. We adapt the AUM to identify ambiguous tasks in crowdsourced learning scenarios, introducing the Weighted AUM (WAUM). The WAUM is an average of AUMs weighted by worker and task dependent scores. We show that the WAUM can help discarding ambiguous tasks from the training set, leading to better generalization or calibration performance. We report improvements with respect to feature-blind aggregation strategies both for simulated settings and for the CIFAR-10H crowdsourced dataset.

High-Dimensional Private Empirical Risk Minimization by Greedy Coordinate Descent

In this paper, we study differentially private empirical risk minimization (DP-ERM). It has been shown that the (worst-case) utility of DP-ERM reduces as the dimension increases. This is a major obstacle to privately learning large machine learning models. In high dimension, it is common for some model's parameters to carry more information than others. To exploit this, we propose a differentially private greedy coordinate descent (DP-GCD) algorithm. At each iteration, DP-GCD privately performs a coordinate-wise gradient step along the gradients' (approximately) greatest entry. We show theoretically that DP-GCD can improve utility by exploiting structural properties of the problem's solution (such as sparsity or quasi-sparsity), with very fast progress in early iterations. We then illustrate this numerically, both on synthetic and real datasets. Finally, we describe promising directions for future work.

Benchopt: Reproducible, efficient and collaborative optimization benchmarks

Numerical validation is at the core of machine learning research as it allows to assess the actual impact of new methods, and to confirm the agreement between theory and practice. Yet, the rapid development of the field poses several challenges: researchers are confronted with a profusion of methods to compare, limited transparency and consensus on best practices, as well as tedious re-implementation work. As a result, validation is often very partial, which can lead to wrong conclusions that slow down the progress of research. We propose Benchopt, a collaborative framework to automate, reproduce and publish optimization benchmarks in machine learning across programming languages and hardware architectures. Benchopt simplifies benchmarking for the community by providing an off-the-shelf tool for running, sharing and extending experiments. To demonstrate its broad usability, we showcase benchmarks on three standard learning tasks: $\ell_2$-regularized logistic regression, Lasso, and ResNet18 training for image classification. These benchmarks highlight key practical findings that give a more nuanced view of the state-of-the-art for these problems, showing that for practical evaluation, the devil is in the details. We hope that Benchopt will foster collaborative work in the community hence improving the reproducibility of research findings.

Stochastic smoothing of the top-K calibrated hinge loss for deep imbalanced classification

In modern classification tasks, the number of labels is getting larger and larger, as is the size of the datasets encountered in practice. As the number of classes increases, class ambiguity and class imbalance become more and more problematic to achieve high top-1 accuracy. Meanwhile, Top-K metrics (metrics allowing K guesses) have become popular, especially for performance reporting. Yet, proposing top-K losses tailored for deep learning remains a challenge, both theoretically and practically. In this paper we introduce a stochastic top-K hinge loss inspired by recent developments on top-K calibrated losses. Our proposal is based on the smoothing of the top-K operator building on the flexible "perturbed optimizer" framework. We show that our loss function performs very well in the case of balanced datasets, while benefiting from a significantly lower computational time than the state-of-the-art top-K loss function. In addition, we propose a simple variant of our loss for the imbalanced case. Experiments on a heavy-tailed dataset show that our loss function significantly outperforms other baseline loss functions.

Supervised learning of analysis-sparsity priors with automatic differentiation

Sparsity priors are commonly used in denoising and image reconstruction. For analysis-type priors, a dictionary defines a representation of signals that is likely to be sparse. In most situations, this dictionary is not known, and is to be recovered from pairs of ground-truth signals and measurements, by minimizing the reconstruction error. This defines a hierarchical optimization problem, which can be cast as a bi-level optimization. Yet, this problem is unsolvable, as reconstructions and their derivative wrt the dictionary have no closed-form expression. However, reconstructions can be iteratively computed using the Forward-Backward splitting (FB) algorithm. In this paper, we approximate reconstructions by the output of the aforementioned FB algorithm. Then, we leverage automatic differentiation to evaluate the gradient of this output wrt the dictionary, which we learn with projected gradient descent. Experiments show that our algorithm successfully learns the 1D Total Variation (TV) dictionary from piecewise constant signals. For the same case study, we propose to constrain our search to dictionaries of 0-centered columns, which removes undesired local minima and improves numerical stability.

LassoBench: A High-Dimensional Hyperparameter Optimization Benchmark Suite for Lasso

Even though Weighted Lasso regression has appealing statistical guarantees, it is typically avoided due to its complex search space described with thousands of hyperparameters. On the other hand, the latest progress with high-dimensional HPO methods for black-box functions demonstrates that high-dimensional applications can indeed be efficiently optimized. Despite this initial success, the high-dimensional HPO approaches are typically applied to synthetic problems with a moderate number of dimensions which limits its impact in scientific and engineering applications. To address this limitation, we propose LassoBench, a new benchmark suite tailored for an important open research topic in the Lasso community that is Weighted Lasso regression. LassoBench consists of benchmarks on both well-controlled synthetic setups (number of samples, SNR, ambient and effective dimensionalities, and multiple fidelities) and real-world datasets, which enable the use of many flavors of HPO algorithms to be improved and extended to the high-dimensional setting. We evaluate 5 state-of-the-art HPO methods and 3 baselines, and demonstrate that Bayesian optimization, in particular, can improve over the methods commonly used for sparse regression while highlighting limitations of these frameworks in very high-dimensions. Remarkably, Bayesian optimization improve the Lasso baselines on 60, 100, 300, and 1000 dimensional problems by 45.7%, 19.2%, 19.7% and 15.5%, respectively.

Differentially Private Coordinate Descent for Composite Empirical Risk Minimization

Machine learning models can leak information about the data used to train them. Differentially Private (DP) variants of optimization algorithms like Stochastic Gradient Descent (DP-SGD) have been designed to mitigate this, inducing a trade-off between privacy and utility. In this paper, we propose a new method for composite Differentially Private Empirical Risk Minimization (DP-ERM): Differentially Private proximal Coordinate Descent (DP-CD). We analyze its utility through a novel theoretical analysis of inexact coordinate descent, and highlight some regimes where DP-CD outperforms DP-SGD, thanks to the possibility of using larger step sizes. We also prove new lower bounds for composite DP-ERM under coordinate-wise regularity assumptions, that are, in some settings, nearly matched by our algorithm. In practical implementations, the coordinate-wise nature of DP-CD updates demands special care in choosing the clipping thresholds used to bound individual contributions to the gradients. A natural parameterization of these thresholds emerges from our theory, limiting the addition of unnecessarily large noise without requiring coordinate-wise hyperparameter tuning or extra computational cost.

Score-Based Change Detection for Gradient-Based Learning Machines

The widespread use of machine learning algorithms calls for automatic change detection algorithms to monitor their behavior over time. As a machine learning algorithm learns from a continuous, possibly evolving, stream of data, it is desirable and often critical to supplement it with a companion change detection algorithm to facilitate its monitoring and control. We present a generic score-based change detection method that can detect a change in any number of components of a machine learning model trained via empirical risk minimization. This proposed statistical hypothesis test can be readily implemented for such models designed within a differentiable programming framework. We establish the consistency of the hypothesis test and show how to calibrate it to achieve a prescribed false alarm rate. We illustrate the versatility of the approach on synthetic and real data.

Spatially relaxed inference on high-dimensional linear models

We consider the inference problem for high-dimensional linear models, when covariates have an underlying spatial organization reflected in their correlation. A typical example of such a setting is high-resolution imaging, in which neighboring pixels are usually very similar. Accurate point and confidence intervals estimation is not possible in this context with many more covariates than samples, furthermore with high correlation between covariates. This calls for a reformulation of the statistical inference problem, that takes into account the underlying spatial structure: if covariates are locally correlated, it is acceptable to detect them up to a given spatial uncertainty. We thus propose to rely on the $\delta$-FWER, that is the probability of making a false discovery at a distance greater than $\delta$ from any true positive. With this target measure in mind, we study the properties of ensembled clustered inference algorithms which combine three techniques: spatially constrained clustering, statistical inference, and ensembling to aggregate several clustered inference solutions. We show that ensembled clustered inference algorithms control the $\delta$-FWER under standard assumptions for $\delta$ equal to the largest cluster diameter. We complement the theoretical analysis with empirical results, demonstrating accurate $\delta$-FWER control and decent power achieved by such inference algorithms.