Borelli@Saints-Pères : M. Ait Saada & S. Perochon

Mira Ait Saada and Sam Perochon present their work at the Borelli@Saints-Pères seminar.

Speaker 1: Mira Ait Saada

Title: Unsupervised Learning from Textual Data with Neural Text Representations

Abstract: Over the last decade, the field of word embeddings has experienced unprecedented interest, particularly with the advent of Transformer models, with a multitude of methods being developed to address all sorts of problems in natural language processing such as named entity recognition, sentiment classification, and question answering systems. All of these objectives are referred to as supervised tasks and require manual labeling to train learning models. In this thesis, however, we consider an unsupervised context where we assume that no prior information about the data is available. Among the most important unsupervised tasks are cluster analysis (or clustering), anomaly detection, and data visualization. Despite their potential, contextual embedding models based on Transformers have been largely overlooked in the unsupervised world of machine learning. This presentation will expose our recent works dedicated to extending the knowledge about the black-box Transformers with a particular focus on unsupervised learning tasks.

Speaker 2: Sam Perochon

Title: Towards an ecological characterization of dysexecutive syndroms: unsupervised action segmentation of untrimmed egocentric videos.

Abstract: Executive functions are the mental processes that help us plan, organize, initiate, and carry out tasks. These functions are necessary for many aspects of daily life, such as problem solving, decision making, and goal directed behavior. Dysexecutive syndrome affects the functioning of executive functions, and are closely related to difficulties in performing activities of daily living [1], influences mental health, physical health, quality of life, school readiness, academic success, career success, marital life and public safety [2]. Dysexecutive syndroms are usually assessed via standardized tests, typically administered in a controlled and standardized environment, such as a clinic or laboratory, and use standardized procedures and materials. However, these tests may not fully reflect an individual's real-life functioning, and may not capture the complexity and variability of dysexecutive symptoms in daily life.Therefore, ecological tests are developed to better assess individual's executive functions in real-life settings, as opposed to in a laboratory or clinical setting.

The Centre Borelli in collaboration with the Hopital d’Instruction des Armées de Percy has developed a project, referred as the SmartFlat project, which aims at enhancing a classical ecological cooking task [3], by leveraging the use of video camera and occulometor. The objective is to make use of the collected multimodal data for the behavioral quantification of dysexecutive syndroms.

In this presentation, we will present our methodology for addressing what we considered the first natural step in analyzing participant behavior: temporal segmentation of video streams.The different parts of the presentation consist in:

1. The detection of background (or outliers) segments, which can help to identify isolated actions and denoised the segments clustering.
2. The change-point detection framework used to detect meaningful action boundaries.
3. Different approaches for the characterization of the participants' behavior in relation to their clinical characteristics.

Bibliography:

• [1] Godefroy, O. (2003). Frontal syndrome and disorders of executive functions. Journal of Neurology, 250(1), 1–6.
• [2] Diamond, A. (2013). Executive Functions. Annual Review of Psychology, 64(1), 135–168.
• [3] Chevignard, M. P., Taillefer, C., Picq, C., Poncet, F., Noulhiane, M., & Pradat-Diehl, P. (2008). Ecological assessment of the dysexecutive syndrome using execution of a cooking task. Neuropsychological Rehabilitation, 18(4), 461–485.