For classification models based on neural networks, the maximum predicted class probability is often used as a confidence score. This score rarely predicts well the probability of making a correct prediction and requires a post-processing calibration step. However, many confidence calibration methods fail for problems with many classes. To address this issue, we transform the problem of calibrating a multiclass classifier into calibrating a single surrogate binary classifier. This approach allows for more efficient use of standard calibration methods. We evaluate our approach on numerous neural networks used for image or text classification and show that it significantly enhances existing calibration methods.

Identifying sources of uncertainty in an image classifier is a crucial challenge. Indeed, the decision process of those models is opaque and does not necessarily correspond to what we might expect. To help characterize classifiers, generative models can be used as they allow the control of visual attributes. Here we use a generative adversarial network to generate images corresponding to how a classifier sees the image. More specifically, we consider the classifier maximum softmax probability as an uncertainty estimation and use it as an additional input to condition the generative model. This allows us to generate images that result in uncertain predictions, giving us a global view of which images are harder to classify. We can also increase the uncertainty of a given image and observe the impact of an attribute, providing a more local understanding of the decision process. We perform experiments on the MNIST dataset, augmented with corruptions. We believe that generative models are a helpful tool to explain the behavior and uncertainties of image classifiers.

Image classifiers should be used with caution in the real world. Performance evaluated on a validation set may not reflect performance in the real world. In particular, classifiers may perform well for conditions that are frequently encountered during training, but poorly for other infrequent conditions. In this study, we hypothesize that recent advances in text-to-image generative models make them valuable for benchmarking computer vision models such as image classifiers: they can generate images conditioned by textual prompts that cause classifier failures, allowing failure conditions to be described with textual attributes. However, their generation cost becomes an issue when a large number of synthetic images need to be generated, which is the case when many different attribute combinations need to be tested. We propose an image classifier benchmarking method as an iterative process that alternates image generation, classifier evaluation, and attribute selection. This method efficiently explores the attributes that ultimately lead to poor behavior detection.

Although deep networks have shown vulnerability to evasion attacks, such attacks have usually unrealistic requirements. Recent literature discussed the possibility to remove or not some of these requirements. This paper contributes to this literature by introducing a carpet-bombing patch attack which has almost no requirement. Targeting the feature representations, this patch attack does not require knowing the network task. This attack decreases accuracy on Imagenet, mAP on Pascal Voc, and IoU on Cityscapes without being aware that the underlying tasks involved classification, detection or semantic segmentation, respectively. Beyond the potential safety issues raised by this attack, the impact of the carpet-bombing attack highlights some interesting property of deep network layer dynamic.

This paper introduces a novel method for the representation of images that is semantic by nature, addressing the question of computation intelligibility in computer vision tasks. More specifically, our proposition is to introduce what we call a semantic bottleneck in the processing pipeline, which is a crossing point in which the representation of the image is entirely expressed with natural language , while retaining the efficiency of numerical representations. We show that our approach is able to generate semantic representations that give state-of-the-art results on semantic content-based image retrieval and also perform very well on image classification tasks. Intelligibility is evaluated through user centered experiments for failure detection.

Zero-Shot learning has been shown to be an efficient strategy for domain adaptation. In this context, this paper builds on the recent work of Bucher et al. [1], which proposed an approach to solve Zero-Shot classification problems (ZSC) by introducing a novel metric learning based objective function. This objective function allows to learn an optimal embedding of the attributes jointly with a measure of similarity between images and attributes. This paper extends their approach by proposing several schemes to control the generation of the negative pairs, resulting in a significant improvement of the performance and giving above state-of-the-art results on three challenging ZSC datasets.