Umeå University's logo

Pretrained language models are an integral part of AI applications, but their high computational cost for training limits accessibility. Initiatives such as BLOOM and STARCODER aim to democratize access to pretrained models for collaborative community development. Despite these efforts, such models encounter challenges such as limited multilingual capabilities, risks of catastrophic forgetting during continual pretraining, and the high costs of training models from scratch, alongside the need to align with AI safety standards and regulatory frameworks. This paper presents AURORA-M, a 15B parameter multilingual open-source model trained on English, Finnish, Hindi, Japanese, Vietnamese, and code. Continually pretrained from STARCODERPLUS on 435B additional tokens, AURORA-M surpasses 2T tokens in total training token count. It is the first open-source multilingual model fine-tuned on human-reviewed safety instructions, thus aligning its development not only with conventional red-teaming considerations, but also with the specific concerns articulated in the Biden-Harris Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. We evaluate AURORA-M across a wide range of tasks and languages, showcasing its robustness against catastrophic forgetting and its superior performance in multilingual settings, particularly in safety evaluations.

The Visual Object Tracking Segmentation VOTS2024 challenge is the twelfth annual tracker benchmarking activity of the VOT initiative. This challenge consolidates the new tracking setup proposed in VOTS2023, which merges short-term and long-term as well as single-target and multiple-target tracking with segmentation masks as the only target location specification. Two sub-challenges are considered. The VOTS2024 standard challenge, focusing on classical objects and the VOTSt2024, which considers objects undergoing a topological transformation. Both challenges use the same performance evaluation methodology. Results of 28 submissions are presented and analyzed. A leaderboard, with participating trackers details, the source code, the datasets, and the evaluation kit are publicly available on the website (https://www.votchallenge.net/vots2024/).

Accurate forecasting and analysis of emerging pandemics play a crucial role in effective public health management and decision-making. Traditional approaches primarily rely on epidemiological data, overlooking other valuable sources of information that could act as sensors or indicators of pandemic patterns. In this paper, we propose a novel framework, MGLEP, that integrates temporal graph neural networks and multi-modal data for learning and forecasting. We incorporate big data sources, including social media content, by utilizing specific pre-trained language models and discovering the underlying graph structure among users. This integration provides rich indicators of pandemic dynamics through learning with temporal graph neural networks. Extensive experiments demonstrate the effectiveness of our framework in pandemic forecasting and analysis, outperforming baseline methods across different areas, pandemic situations, and prediction horizons. The fusion of temporal graph learning and multi-modal data enables a comprehensive understanding of the pandemic landscape with less time lag, cheap cost, and more potential information indicators.

Real-world sound signals exhibit various aspects of grouping and profiling behaviors, such as being recorded from identical sources, having similar environmental settings, or encountering related background noises. In this work, we propose novel neural profiling networks (NeuProNet) capable of learning and extracting high-level unique profile representations from sounds. An end-to-end framework is developed so that any backbone architectures can be plugged in and trained, achieving better performance in any downstream sound classification tasks. We introduce an in-batch profile grouping mechanism based on profile awareness and attention pooling to produce reliable and robust features with contrastive learning. Furthermore, extensive experiments are conducted on multiple benchmark datasets and tasks to show that neural computing models under the guidance of our framework gain significant performance gaps across all evaluation tasks. Particularly, the integration of NeuProNet surpasses recent state-of-the-art (SoTA) approaches on UrbanSound8K and VocalSound datasets with statistically significant improvements in benchmarking metrics, up to 5.92% in accuracy compared to the previous SoTA method and up to 20.19% compared to baselines. Our work provides a strong foundation for utilizing neural profiling for machine learning tasks.

Linguistic data, a component critical not only for research in a variety of fields but also for the development of various Natural Language Processing (NLP) applications, can contain personal information. As a result, its accessibility is limited, both from a legal and an ethical standpoint. One of the solutions is the pseudonymization of the data. Key stages of this process include the identification of sensitive elements and the generation of suitable surrogates in a way that the data is still useful for the intended task. Within this paper, we conduct an analysis of tagsets that have previously been utilized in anonymization and pseudonymization. We also investigate what kinds of Personally Identifiable Information (PII) appear in various domains. These reveal that none of the analyzed tagsets account for all of the PII types present cross-domain at the level of detailedness seemingly required for pseudonymization. We advocate for a universal system of tags for categorizing PIIs leading up to their replacement. Such categorization could facilitate the generation of grammatically, semantically, and sociolinguistically appropriate surrogates for the kinds of information that are considered sensitive in a given domain, resulting in a system that would enable dynamic pseudonymization while keeping the texts readable and useful for future research in various fields.

This paper investigates a novel graph-based representation of sound waves inspired by the physical phenomenon of correlated vibrations. We propose a Wave2Graph framework for integrating multiple acoustic representations, including the spectrum of frequencies and correlations, into various neural computing architectures to achieve new state-of-the-art performances in sound classification. The capability and reliability of our end-to-end framework are evidently demonstrated in voice pathology for low-cost and non-invasive mass-screening of medical conditions, including respiratory illnesses and Alzheimer's Dementia. We conduct extensive experiments on multiple public benchmark datasets (ICBHI and ADReSSo) and our real-world dataset (IJSound: Respiratory disease detection using coughs and breaths). Wave2Graph framework consistently outperforms previous state-of-the-art methods with a large magnitude, up to 7.65% improvement, promising the usefulness of graph-based representation in signal processing and machine learning.

We introduce ADCluster, a deep document clustering approach based on language models that is trained to adapt to the clustering task. This adaptability is achieved through an iterative process where K-Means clustering is applied to the dataset, followed by iteratively training a deep classifier with generated pseudo-labels – an approach referred to as inner adaptation. The model is also able to adapt to changes in the data as new documents are added to the document collection. The latter type of adaptation, outer adaptation, is obtained by resuming the inner adaptation when a new chunk of documents has arrived. We explore two outer adaptation strategies, namely accumulative adaptation (training is resumed on the accumulated set of all documents) and non-accumulative adaptation (training is resumed using only the new chunk of data). We show that ADCluster outperforms established document clustering techniques on medium and long-text documents by a large margin. Additionally, our approach outperforms well-established baseline methods under both the accumulative and non-accumulative outer adaptation scenarios.

This paper firstly proposes a simple yet efficient generalized approach to apply differential privacy to text representation (i.e., word embedding). Based on it, we propose a user-level approach to learn personalized differentially private word embedding model on user generated contents (UGC). To our best knowledge, this is the first work of learning user-level differentially private word embedding model from text for sharing. The proposed approaches protect the privacy of the individual from re-identification, especially provide better trade-off of privacy and data utility on UGC data for sharing. The experimental results show that the trained embedding models are applicable for the classic text analysis tasks (e.g., regression). Moreover, the proposed approaches of learning differentially private embedding models are both framework- and dataindependent, which facilitates the deployment and sharing. The source code is available at https://github.com/sonvx/dpText.

Accessibility of research data is critical for advances in many research fields, but textual data often cannot be shared due to the personal and sensitive information which it contains, e.g names or political opinions. General Data Protection Regulation (GDPR) suggests pseudonymization as a solution to secure open access to research data, but we need to learn more about pseudonymization as an approach before adopting it for manipulation of research data. This paper outlines a research agenda within pseudonymization, namely need of studies into the effects of pseudonymization on unstructured data in relation to e.g. readability and language assessment, as well as the effectiveness of pseudonymization as a way of protecting writer identity, while also exploring different ways of developing context-sensitive algorithms for detection, labelling and replacement of personal information in unstructured data. The recently granted project on pseudonymization 'Grandma Karl is 27 years old'¹ addresses exactly those challenges. ¹.https://spraakbanken.gu.se/en/projects/mormor-karl