Objective non-stationarity measures are resource intensive and impose critical limitations for real-time processing solutions. In this paper, a novel Hard Label Criteria (HLC) algorithm is proposed to generate global non-stationarity labels for acoustic signals, enabling supervised learning strategies to be trained as stationarity estimators. The HLC is first evaluated on state-of-the-art general-purpose acoustic models, demonstrating that these models capture stationarity information. Furthermore, the first-of-its-kind HLC-based Network for Acoustic Non-Stationarity Assessment (NANSA) is proposed. NANSA models outperform competing approaches, achieving up to 99% classification accuracy, while solving the computational infeasibility of traditional objective measures.
Automatic Text Summarization (ATS) is a Natural Language Processing (NLP) task essential for handling large volumes of information. ATS can be classified into two main types: extractive and abstractive. Extractive summarization selects sentences or phrases directly from the source text(s), while abstractive summarization generates new sentences that try to capture the original meaning of the source text(s). This paper describes our efforts to perform extractive single-document summarization in multilingual contexts. Although various summarization methods, such as PreSumm and HiStruct+, have shown promising results on English corpora like CNN/DM, there is a significant gap in applying these methods to other languages, especially Brazilian Portuguese. Additionally, these summarizers were evaluated with traditional metrics like ROUGE, which has limitations as it primarily measures superficial text overlap. To fill these gaps, we evaluate the effectiveness of these state-of-the-art methods on the CSTNews corpus (with news texts in Brazilian Portuguese) employing ROUGE and the recent BLANC metric, which measures how much the generated summary aids a pre-trained language model (like BERT) in understanding the document. Our contributions include the results and comparison of adapted models, the discussion of the BLANC metric in contrast to ROUGE, and the expansion of resources available to the Portuguese and multilingual NLP community.
While hospitals routinely gather patient data, such as X-ray images, the challenge of sharing this data across multiple institutions to create a comprehensive and large dataset is hampered by privacy concerns. Consequently, this limitation affects the effectiveness of state-of-the-art deep neural networks for tasks like identifying lung diseases in medical images, as they require substantial annotated data. Federated Learning offers a solution by enabling collaborative training across multiple edge devices or sites, where updates (e.g., neural network weights) are aggregated without sharing patient data, thus maintaining privacy. This work introduces a federated-learning-based approach for automatically detecting lung diseases in chest X-ray images, focusing on preserving data privacy and enhancing robustness. Our approach follows the federated learning protocol: decentralized training of neural networks on data from multiple sites (hospitals) and centralized aggregation of knowledge in the server. The solution presents promising results in identifying fourteen lung diseases compared to three baselines within a simulated environment comprising chest X-ray images from five distinct sites.
Accurate analysis of human movement is essential for developing effective human-robot interaction (HRI) interfaces. Vision-based pose estimation tools like MediaPipe, which operate markerlessly with a single camera, offer a low-cost alternative but lack extensive validation across different capture conditions. This study evaluates the consistency of MediaPipe's pose estimations during three basic human movements (squatting, pulling, pushing), captured from three camera angles (0 degrees, 45 degrees, and 90 degrees). Joint angles were analyzed, and Pearson's correlations were computed between time series across views. Results reveal how MediaPipe's accuracy varies with movement and camera placement, offering insights into its applicability for HRI and human movement analysis.
Modern approaches for keyword spotting rely on training deep neural networks on large static datasets with i.i.d. distributions. However, the resulting models tend to underperform when presented with changing data regimes in real-life applications. This work investigates a simple but effective online continual learning method that updates a keyword spotter on-device via SGD as new data becomes available. Contrary to previous research, this work focuses on learning the same KWS task, which covers most commercial applications. During experiments with dynamic audio streams in different scenarios, that method improves the performance of a pre-trained small-footprint model by 34%. Moreover, experiments demonstrate that, compared to a naive online learning implementation, conditional model updates based on its performance in a small hold-out set drawn from the training distribution mitigate catastrophic forgetting.
