Visual Document Retrieval (VDR) models mostly rely on late interaction architectures, in which documents are represented by a set of local patch embeddings and then matched against query tokens. While efficient, this architecture prioritizes local similarity over global layout structure of documents to estimate relevancy between documents and query. In practice, this leads to errors as relevance originates from layout structure of documents with heterogeneous layouts combining figures, tables, and text. We make document layout learnable without changing inference. We propose a multimodal encoder that augments local patch representations with a global layout embedding, trained via textual descriptions encoding document layout information. Across four ViDoRe-v2 datasets, our model improves over the strongest architecturally comparable ColPali/ColQwen baseline by +2.4 nDCG@5 and +2.3 MAP@5, with statistically significant per-dataset gains over ColQwen.

This study presents a novel visualization approach to explainable artificial intelligence for graph-based visual question answering (VQA) systems. The method focuses on identifying false answer predictions by the model and offers users the opportunity to directly correct mistakes in the input space, thus facilitating dataset curation. The decision-making process of the model is demonstrated by highlighting certain internal states of a graph neural network (GNN). The proposed system is built on top of a GraphVQA framework that implements various GNN-based models for VQA trained on the GQA dataset. The authors evaluated their tool through the demonstration of identified use cases, quantitative measures, and a user study conducted with experts from machine learning, visualization, and natural language processing domains. The authors' findings highlight the prominence of their implemented features in supporting the users with incorrect prediction identification and identifying the underlying issues. Additionally, their approach is easily extendable to similar models aiming at graph-based question answering.

Dual encoder architectures like Clip models map two types of inputs into a shared embedding space and predict similarities between them. Despite their wide application, it is, however, not understood how these models compare their two inputs. Common first-order feature-attribution methods explain importances of individual features and can, thus, only provide limited insights into dual encoders, whose predictions depend on interactions between features. In this paper, we first derive a second-order method enabling the attribution of predictions by any differentiable dual encoder onto feature-interactions between its inputs. Second, we apply our method to Clip models and show that they learn fine-grained correspondences between parts of captions and regions in images. They match objects across input modes and also account for mismatches. This intrinsic visual-linguistic grounding ability, however, varies heavily between object classes, exhibits pronounced out-of-domain effects and we can identify individual errors as well as systematic failure categories. Code is publicly available: https://github.com/lucasmllr/exCLIP

The large success of deep learning based methods in Visual Question Answering (VQA) has concurrently increased the demand for explainable methods. Most methods in Explainable Artificial Intelligence (XAI) focus on generating post-hoc explanations rather than taking an intrinsic approach, the latter characterizing an interpretable model. In this work, we introduce an interpretable approach for graph-based VQA and demonstrate competitive performance on the GQA dataset. This approach bridges the gap between interpretability and performance. Our model is designed to intrinsically produce a subgraph during the question-answering process as its explanation, providing insight into the decision making. To evaluate the quality of these generated subgraphs, we compare them against established post-hoc explainability methods for graph neural networks, and perform a human evaluation. Moreover, we present quantitative metrics that correlate with the evaluations of human assessors, acting as automatic metrics for the generated explanatory subgraphs. Our implementation is available at https://github.com/DigitalPhonetics/Intrinsic-Subgraph-Generation-for-VQA.

Explainable artificial intelligence (XAI) aims to make machine learning models more transparent. While many approaches focus on generating explanations post-hoc, interpretable approaches, which generate the explanations intrinsically alongside the predictions, are relatively rare. In this work, we integrate different discrete subset sampling methods into a graph-based visual question answering system to compare their effectiveness in generating interpretable explanatory subgraphs intrinsically. We evaluate the methods on the GQA dataset and show that the integrated methods effectively mitigate the performance trade-off between interpretability and answer accuracy, while also achieving strong co-occurrences between answer and question tokens. Furthermore, we conduct a human evaluation to assess the interpretability of the generated subgraphs using a comparative setting with the extended Bradley-Terry model, showing that the answer and question token co-occurrence metrics strongly correlate with human preferences. Our source code is publicly available.

Scene-graph-based Visual Question Answering (VQA) has emerged as a burgeoning field in Deep Learning research, with a growing demand for robust and interpretable VQA systems. In this paper, we present a novel visual analysis approach that addresses two critical objectives in VQA: identifying and correcting prediction issues and providing insights into model decision-making processes through visualizing internal information. Our approach builds on the GraphVQA framework, which uses graph neural networks to process scene graphs representing images and which was trained on the widely-used GQA dataset. Our analysis tool aims at users familiar with the basics of graph-based VQA. By leveraging query-based scene analysis and visualization of crucial internal states, we are able to detect and pinpoint reasons for inaccurate predictions, facilitating model refinement and dataset curation. Identifying expressive internal states is a challenge. Through rigorous computer-based evaluations and presentation of a use case, we demonstrate the effectiveness of our analysis tool and model state visualization.

Prosody is closely linked to the identity of a speaker, leading to individual pitch and intonation patterns. Therefore, it is challenging in speaker anonymization to generate speech utterances that both keep the original audio's main prosodic structure and preserve the speaker's privacy. In this paper, we present a system that extends a speech-to-text-to-speech anonymization pipeline with prosody cloning and show how to control the cloning by multiplying pitch and energy sequences with random offset values. Using automatic and human evaluation, we find this combination to successfully overcome the privacy-utility trade-off for prosody by achieving high privacy and high pitch correlation scores. At the same time, the anonymized utterances prove to reproduce the original voice distinctiveness and content with high intelligibility and only a small loss in naturalness, making them suitable for downstream applications.

Image-Text-Matching (ITM) is one of the defacto methods of learning generalized representations from a large corpus in Vision and Language (VL). However, due to the weak association between the web-collected image-text pairs, models fail to show a fine-grained understanding of the combined semantics of these modalities. To address this issue we propose Hard Negative Captions (HNC): an automatically created dataset containing foiled hard negative captions for ITM training towards achieving fine-grained cross-modal comprehension in VL. Additionally, we provide a challenging manually-created test set for benchmarking models on a fine-grained cross-modal mismatch task with varying levels of compositional complexity. Our results show the effectiveness of training on HNC by improving the models' zero-shot capabilities in detecting mismatches on diagnostic tasks and performing robustly under noisy visual input scenarios. Also, we demonstrate that HNC models yield a comparable or better initialization for fine-tuning

Customizing voice and speaking style in a speech synthesis system with intuitive and fine-grained controls is challenging, given that little data with appropriate labels is available.Furthermore, editing an existing human's voice also comes with ethical concerns.In this paper, we propose a method to generate artificial speaker embeddings that cannot be linked to a real human while offering intuitive and fine-grained control over the voice and speaking style of the embeddings, without requiring any labels for speaker or style.The artificial and controllable embeddings can be fed to a speech synthesis system, conditioned on embeddings of real humans during training, without sacrificing privacy during inference.

In order to protect the privacy of speech data, speaker anonymization aims for hiding the identity of a speaker by changing the voice in speech recordings. This typically comes with a privacy-utility trade-off between protection of individuals and usability of the data for downstream applications. One of the challenges in this context is to create non-existent voices that sound as natural as possible. In this work, we propose to tackle this issue by generating speaker embeddings using a generative adversarial network with Wasserstein distance as cost function. By incorporating these artificial embeddings into a speech-to-text-to-speech pipeline, we outperform previous approaches in terms of privacy and utility. According to standard objective metrics and human evaluation, our approach generates intelligible and content-preserving yet privacy-protecting versions of the original recordings.

In this work, we propose a speaker anonymization pipeline that leverages high quality automatic speech recognition and synthesis systems to generate speech conditioned on phonetic transcriptions and anonymized speaker embeddings. Using phones as the intermediate representation ensures near complete elimination of speaker identity information from the input while preserving the original phonetic content as much as possible. Our experimental results on LibriSpeech and VCTK corpora reveal two key findings: 1) although automatic speech recognition produces imperfect transcriptions, our neural speech synthesis system can handle such errors, making our system feasible and robust, and 2) combining speaker embeddings from different resources is beneficial and their appropriate normalization is crucial. Overall, our final best system outperforms significantly the baselines provided in the Voice Privacy Challenge 2020 in terms of privacy robustness against a lazy-informed attacker while maintaining high intelligibility and naturalness of the anonymized speech.

Speaker anonymization is the task of modifying speech recordings to hide the identity of the original speaker by changing the voice in the audio. Simultaneously, the anonymized audio should remain usable for downstream tasks and thus keep other information of the original audio like the linguistic content. This typically creates a privacy-utility trade-off of anonymization techniques. In our submission to the VoicePrivacy 2022 Challenge, we aim to reduce this trade-off by creating a speech-to-speech pipeline that (a) eliminates all clues about speaker identity by reducing the audio to phonetic transcriptions, (b) generates a new, non-existent voice using a Generative Adversarial Network, leading to artificial yet natural-like and distinctive speakers, and (c) synthesizes an anonymous version of the original utterance based on the transcriptions, anonymous speaker embedding, and estimated pitch. According to the ob-jective evaluation, this anonymization method leads to almost perfect privacy and voice distinctiveness, and clearly outperforms all baseline systems for these two metrics. For the speech recognition utility metric, we achieve similar good results on LibriSpeech and much better ones on VCTK as compared to the baselines and the original non-anonymized data. Solely for pitch correlation, we only just meet the required threshold because our system does not use the original pitch trajectory for synthesis. Overall, our approach successfully hides the speaker identity while keeping the linguistic content, proving to be generally more effective than any of the baselines of the VoicePri-vacy 2022 Challenge.

On the way towards general Visual Question Answering (VQA) systems that are able to answer arbitrary questions, the need arises for evaluation beyond single-metric leaderboards for specific datasets. To this end, we propose a browser-based benchmarking tool for researchers and challenge organizers, with an API for easy integration of new models and datasets to keep up with the fast-changing landscape of VQA. Our tool helps test generalization capabilities of models across multiple datasets, evaluating not just accuracy, but also performance in more realistic real-world scenarios such as robustness to input noise. Additionally, we include metrics that measure biases and uncertainty, to further explain model behavior. Interactive filtering facilitates discovery of problematic behavior, down to the data sample level. As proof of concept, we perform a case study on four models. We find that state-of-the-art VQA models are optimized for specific tasks or datasets, but fail to generalize even to other in-domain test sets, for example they can not recognize text in images. Our metrics allow us to quantify which image and question embeddings provide most robustness to a model. All code s publicly available.