Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by persistent difficulties including repetitive patterns of behavior known as stereotypical motor movements (SMM). So far, several techniques have been implemented to track and identify SMMs. In this context, we propose a deep learning approach for SMM recognition, namely, convolutional neural networks (CNN) in time and frequency-domains. To solve the intrasubject SMM variability, we propose a robust CNN model for SMM detection within subjects, whose parameters are set according to a proper analysis of SMM signals, thereby outperforming state-of-the-art SMM classification works. And, to solve the intersubject variability, we propose a global, fast, and light-weight framework for SMM detection across subjects which combines a knowledge transfer technique with an SVM classifier, therefore resolving the “real-life” medical issue associated with the lack of supervised SMMs per testing subject in particular. We further show that applying transfer learning across domains instead of transfer learning within the same domain also generalizes to the SMM target domain, thus alleviating the problem of the lack of supervised SMMs in general.

Time series classification is an important field in time series data-mining which have covered broad applications so far. Although it has attracted great interests during last decades, it remains a challenging task and falls short of efficiency due to the nature of its data: high dimensionality, large in data size and updating continuously. With the advent of deep learning, new methods have been developed, especially Convolutional Neural Network (CNN) models. In this paper, we present a review of our time series CNN approaches including: (i) a data-level approach based on encoding time series into frequency-domain signals via the Stockwell transform, (ii) an algorithm-level approach based on an adaptive convolutional layer filter that suits the time series in hand, and (iii) another algorithm-level approach adapted to time series classification tasks with limited annotated data, which is a global, fast and lightweight framework based on a transfer learning technique with a source learning task similar or different but related to the target learning task. These approaches are implemented on identifying human activities including normal movements of typical subjects and disorder-related movements such as stereotypical motor movements of autistic subjects. Experimental results show that our approaches improve performance of time series classification.

No abstract

Unlike Latin, the recognition of Phoenician handwritten characters remains at the level of research and experimentation. In fact, such recognition can contribute to performing tasks such as automatic processing of Phoenician administrative records and scripts, the digitization and the safeguarding of the written Phoenician cultural heritage. As such, the availability of a reference database for Phoenician handwritten characters is crucial to carry out these tasks. To this matter, a database for Phoenician handwritten characters (PHCDB) is introduced for the first time in this paper. We also explore the challenges in the recognition of Phoenician handwritten characters by proposing a deep learning architecture trained on our database. Furthermore, we propose a transfer learning system based on Phoenician character shapes to improve the recognition performance of Tifinagh Handwritten character, and we thereby affirm the possibility of the Tifinagh alphabet being derived from the Phoenician alphabet. Finally, based on Phoenician characters, we introduce a fast, global and light-weight transfer learning system for the recognition of any alphabet which lacks annotated data.

Many real-world data mining applications involve obtaining predictive models using imbalanced datasets. Frequently, the least common target variables present within datasets are associated with events that are highly relevant for end users. When these variables are nominal, we have a class-imbalance problem which has been thoroughly studied within machine learning. As for regression tasks where target variables are continuous, few predictive models and evaluation techniques exist. This paper proposes a solution to these challenges. First, we introduce a cost-sensitive learning algorithm based on a neural network trained on the minimization of a biased loss function. Results show a higher or comparable performance and convergence speed to existent techniques. Second, we develop new approaches for performance assessment of regression tasks within imbalanced domains by proposing new scalar measures, namely Geometric Mean Error (GME) and Class-Weighted Error (CWE), as well as new graphical-based measures, namely REC TPR , REC TNR , REC G − Mean and REC CWA curves. Unlike standard measures, our evaluation strategies are shown to be more robust to data imbalance as they reflect the performance of both rare and frequent events.