FaceX: Understanding Face Attribute Classifiers through Summary Model Explanations

EXplainable Artificial Intelligence (XAI) approaches are widely applied for identifying fairness issues in Artificial Intelligence (AI) systems. However, in the context of facial analysis, existing XAI approaches, such as pixel attribution methods, offer explanations for individual images, posing challenges in assessing the overall behavior of a model, which would require labor-intensive manual inspection of a very large number of instances and leaving to the human the task of drawing a general impression of the model behavior from the individual outputs. Addressing this limitation, we introduce FaceX, the first method that provides a comprehensive understanding of face attribute classifiers through summary model explanations. Specifically, FaceX leverages the presence of distinct regions across all facial images to compute a region-level aggregation of model activations, allowing for the visualization of the model's region attribution across 19 predefined regions of interest in facial images, such as hair, ears, or skin. Beyond spatial explanations, FaceX enhances interpretability by visualizing specific image patches with the highest impact on the model's decisions for each facial region within a test benchmark. Through extensive evaluation in various experimental setups, including scenarios with or without intentional biases and mitigation efforts on four benchmarks, namely CelebA, FairFace, CelebAMask-HQ, and Racial Faces in the Wild, FaceX demonstrates high effectiveness in identifying the models' biases.

MAVias: Mitigate any Visual Bias

Mitigating biases in computer vision models is an essential step towards the trustworthiness of artificial intelligence models. Existing bias mitigation methods focus on a small set of predefined biases, limiting their applicability in visual datasets where multiple, possibly unknown biases exist. To address this limitation, we introduce MAVias, an open-set bias mitigation approach leveraging foundation models to discover spurious associations between visual attributes and target classes. MAVias first captures a wide variety of visual features in natural language via a foundation image tagging model, and then leverages a large language model to select those visual features defining the target class, resulting in a set of language-coded potential visual biases. We then translate this set of potential biases into vision-language embeddings and introduce an in-processing bias mitigation approach to prevent the model from encoding information related to them. Our experiments on diverse datasets, including CelebA, Waterbirds, ImageNet, and UrbanCars, show that MAVias effectively detects and mitigates a wide range of biases in visual recognition tasks outperforming current state-of-the-art.

BAdd: Bias Mitigation through Bias Addition

Computer vision (CV) datasets often exhibit biases that are perpetuated by deep learning models. While recent efforts aim to mitigate these biases and foster fair representations, they fail in complex real-world scenarios. In particular, existing methods excel in controlled experiments involving benchmarks with single-attribute injected biases, but struggle with multi-attribute biases being present in well-established CV datasets. Here, we introduce BAdd, a simple yet effective method that allows for learning fair representations invariant to the attributes introducing bias by incorporating features representing these attributes into the backbone. BAdd is evaluated on seven benchmarks and exhibits competitive performance, surpassing state-of-the-art methods on both single- and multi-attribute benchmarks. Notably, BAdd achieves +27.5% and +5.5% absolute accuracy improvements on the challenging multi-attribute benchmarks, FB-Biased-MNIST and CelebA, respectively.

CycleMix: Mixing Source Domains for Domain Generalization in Style-Dependent Data

As deep learning-based systems have become an integral part of everyday life, limitations in their generalization ability have begun to emerge. Machine learning algorithms typically rely on the i.i.d. assumption, meaning that their training and validation data are expected to follow the same distribution, which does not necessarily hold in practice. In the case of image classification, one frequent reason that algorithms fail to generalize is that they rely on spurious correlations present in training data, such as associating image styles with target classes. These associations may not be present in the unseen test data, leading to significant degradation of their effectiveness. In this work, we attempt to mitigate this Domain Generalization (DG) problem by training a robust feature extractor which disregards features attributed to image-style but infers based on style-invariant image representations. To achieve this, we train CycleGAN models to learn the different styles present in the training data and randomly mix them together to create samples with novel style attributes to improve generalization. Experimental results on the PACS DG benchmark validate the proposed method.

Sampling Strategies for Mitigating Bias in Face Synthesis Methods

Synthetically generated images can be used to create media content or to complement datasets for training image analysis models. Several methods have recently been proposed for the synthesis of high-fidelity face images; however, the potential biases introduced by such methods have not been sufficiently addressed. This paper examines the bias introduced by the widely popular StyleGAN2 generative model trained on the Flickr Faces HQ dataset and proposes two sampling strategies to balance the representation of selected attributes in the generated face images. We focus on two protected attributes, gender and age, and reveal that biases arise in the distribution of randomly sampled images against very young and very old age groups, as well as against female faces. These biases are also assessed for different image quality levels based on the GIQA score. To mitigate bias, we propose two alternative methods for sampling on selected lines or spheres of the latent space to increase the number of generated samples from the under-represented classes. The experimental results show a decrease in bias against underrepresented groups and a more uniform distribution of the protected features at different levels of image quality.

Effector: A Python package for regional explanations

Global feature effect methods explain a model outputting one plot per feature. The plot shows the average effect of the feature on the output, like the effect of age on the annual income. However, average effects may be misleading when derived from local effects that are heterogeneous, i.e., they significantly deviate from the average. To decrease the heterogeneity, regional effects provide multiple plots per feature, each representing the average effect within a specific subspace. For interpretability, subspaces are defined as hyperrectangles defined by a chain of logical rules, like age's effect on annual income separately for males and females and different levels of professional experience. We introduce Effector, a Python library dedicated to regional feature effects. Effector implements well-established global effect methods, assesses the heterogeneity of each method and, based on that, provides regional effects. Effector automatically detects subspaces where regional effects have reduced heterogeneity. All global and regional effect methods share a common API, facilitating comparisons between them. Moreover, the library's interface is extensible so new methods can be easily added and benchmarked. The library has been thoroughly tested, ships with many tutorials (https://xai-effector.github.io/) and is available under an open-source license at PyPi (https://pypi.org/project/effector/) and Github (https://github.com/givasile/effector).

C-XGBoost: A tree boosting model for causal effect estimation

Causal effect estimation aims at estimating the Average Treatment Effect as well as the Conditional Average Treatment Effect of a treatment to an outcome from the available data. This knowledge is important in many safety-critical domains, where it often needs to be extracted from observational data. In this work, we propose a new causal inference model, named C-XGBoost, for the prediction of potential outcomes. The motivation of our approach is to exploit the superiority of tree-based models for handling tabular data together with the notable property of causal inference neural network-based models to learn representations that are useful for estimating the outcome for both the treatment and non-treatment cases. The proposed model also inherits the considerable advantages of XGBoost model such as efficiently handling features with missing values requiring minimum preprocessing effort, as well as it is equipped with regularization techniques to avoid overfitting/bias. Furthermore, we propose a new loss function for efficiently training the proposed causal inference model. The experimental analysis, which is based on the performance profiles of Dolan and Mor{\'e} as well as on post-hoc and non-parametric statistical tests, provide strong evidence about the effectiveness of the proposed approach.

On the Out-Of-Distribution Robustness of Self-Supervised Representation Learning for Phonocardiogram Signals

Objective: Despite the recent increase in research activity, deep-learning models have not yet been widely accepted in medicine. The shortage of high-quality annotated data often hinders the development of robust and generalizable models, which do not suffer from degraded effectiveness when presented with newly-collected, out-of-distribution (OOD) datasets. Methods: Contrastive Self-Supervised Learning (SSL) offers a potential solution to the scarcity of labeled data as it takes advantage of unlabeled data to increase model effectiveness and robustness. In this research, we propose applying contrastive SSL for detecting abnormalities in phonocardiogram (PCG) samples by learning a generalized representation of the signal. Specifically, we perform an extensive comparative evaluation of a wide range of audio-based augmentations and evaluate trained classifiers on multiple datasets across different downstream tasks. Results: We experimentally demonstrate that, depending on its training distribution, the effectiveness of a fully-supervised model can degrade up to 32% when evaluated on unseen data, while SSL models only lose up to 10% or even improve in some cases. Conclusions: Contrastive SSL pretraining can assist in providing robust classifiers which can generalize to unseen, OOD data, without relying on time- and labor-intensive annotation processes by medical experts. Furthermore, the proposed extensive evaluation protocol sheds light on the most promising and appropriate augmentations for robust PCG signal processing. Significance: We provide researchers and practitioners with a roadmap towards producing robust models for PCG classification, in addition to an open-source codebase for developing novel approaches.

FRCSyn Challenge at WACV 2024:Face Recognition Challenge in the Era of Synthetic Data

Despite the widespread adoption of face recognition technology around the world, and its remarkable performance on current benchmarks, there are still several challenges that must be covered in more detail. This paper offers an overview of the Face Recognition Challenge in the Era of Synthetic Data (FRCSyn) organized at WACV 2024. This is the first international challenge aiming to explore the use of synthetic data in face recognition to address existing limitations in the technology. Specifically, the FRCSyn Challenge targets concerns related to data privacy issues, demographic biases, generalization to unseen scenarios, and performance limitations in challenging scenarios, including significant age disparities between enrollment and testing, pose variations, and occlusions. The results achieved in the FRCSyn Challenge, together with the proposed benchmark, contribute significantly to the application of synthetic data to improve face recognition technology.

Regionally Additive Models: Explainable-by-design models minimizing feature interactions

Generalized Additive Models (GAMs) are widely used explainable-by-design models in various applications. GAMs assume that the output can be represented as a sum of univariate functions, referred to as components. However, this assumption fails in ML problems where the output depends on multiple features simultaneously. In these cases, GAMs fail to capture the interaction terms of the underlying function, leading to subpar accuracy. To (partially) address this issue, we propose Regionally Additive Models (RAMs), a novel class of explainable-by-design models. RAMs identify subregions within the feature space where interactions are minimized. Within these regions, it is more accurate to express the output as a sum of univariate functions (components). Consequently, RAMs fit one component per subregion of each feature instead of one component per feature. This approach yields a more expressive model compared to GAMs while retaining interpretability. The RAM framework consists of three steps. Firstly, we train a black-box model. Secondly, using Regional Effect Plots, we identify subregions where the black-box model exhibits near-local additivity. Lastly, we fit a GAM component for each identified subregion. We validate the effectiveness of RAMs through experiments on both synthetic and real-world datasets. The results confirm that RAMs offer improved expressiveness compared to GAMs while maintaining interpretability.