Seminario de Probabilidad y Estadística

This presentation will focus on computational behavioral phenotyping, illustrating how machine learning (ML) methods can be leveraged to transform raw complex multimodal health data into objective and interpretable biomarkers. I first present computer vision and ML algorithms that (i) extract digital behavioral biomarkers from video and gamified visuomotor assessments and (ii) combine them for early autism screening. I then describe computational approaches for analyzing egocentric video to characterize the execution of a chocolate cake cooking task, by producing compact symbolic representations of behavioral trajectories that enable alignment and comparison across patients with sensorimotor and executive disorders. Finally, I present a methodology for learning compact and generalizable representations of wearable biosignals to improve the sensitivity to subtle cardiovascular health changes.

[1] Perochon S., Di Martino J.M. et al. A scalable computational approach to assessing response to name in toddlers with autism. J Child Psychol Psychiatr (2021).

[2] Perochon S., Di Martino J.M. et al. A tablet-based game for the assessment of visual motor skills in autistic children. npj Digit. Med. (2023).

[3] Perochon, S., Di Martino, J.M. et al. Early detection of autism using digital behavioral phenotyping. Nat Med. (2023).

[4] Perochon S., Oudre L. Recursive prototyping for computational behavior analysis from egocentric videos. In review at ICPR 2026

[5] Perochon, S. et al. Time-varying representations of longitudinal biosignals using self-supervised learning. NeurIPS 2024 Workshop: Self-Supervised Learning-Theory and Practice.

Esta charla aborda un tema con el que me he obsesionado desde hace ya varios años: la extracción de datos presentes en las actas de estado civil, tales como las actas de nacimiento. El problema se sitúa en la intersección de visión por computadora, aprendizaje profundo y la privacidad de datos. En la primera parte, daré contexto exponiendo sobre mi trabajo de tesis de la Maestría en Ciencia de Datos y Aprendizaje Automático (Fing, Udelar), que aborda el problema de transcribir automáticamente actas de nacimiento uruguayas manuscritas. Iremos de pipelines clásicos de segmentación y reconocimiento hasta modelos extremo a extremo, y veremos cómo, con conjuntos de datos pequeños y estrategias de anotación bien elegidas, es posible alcanzar resultados muy precisos. En la segunda parte, les contaré sobre el trabajo que vengo realizando en el marco de mi doctorado en la Université Paris-Saclay, en torno a la pregunta: ¿qué hacemos cuando los datos reales no se pueden utilizar (ni para entrenar, ni para publicar) por restricciones de privacidad? Para responder esto, presentaré un método no supervisado para generar documentos sintéticos realistas a partir de formularios manuscritos, que permite entrenar modelos sin exponer información sensible.

This presentation will focus on computational behavioral phenotyping, illustrating how machine learning (ML) methods can be leveraged to transform raw complex multimodal health data into objective and interpretable biomarkers. I first present computer vision and ML algorithms that (i) extract digital behavioral biomarkers from video and gamified visuomotor assessments and (ii) combine them for early autism screening. I then describe computational approaches for analyzing egocentric video to characterize the execution of a chocolate cake cooking task, by producing compact symbolic representations of behavioral trajectories that enable alignment and comparison across patients with sensorimotor and executive disorders. Finally, I present a methodology for learning compact and generalizable representations of wearable biosignals to improve the sensitivity to subtle cardiovascular health changes.

[1] Perochon S., Di Martino J.M. et al. A scalable computational approach to assessing response to name in toddlers with autism. J Child Psychol Psychiatr (2021).

[2] Perochon S., Di Martino J.M. et al. A tablet-based game for the assessment of visual motor skills in autistic children. npj Digit. Med. (2023).

[3] Perochon, S., Di Martino, J.M. et al. Early detection of autism using digital behavioral phenotyping. Nat Med. (2023).

[4] Perochon S., Oudre L. Recursive prototyping for computational behavior analysis from egocentric videos. In review at ICPR 2026

[5] Perochon, S. et al. Time-varying representations of longitudinal biosignals using self-supervised learning. NeurIPS 2024 Workshop: Self-Supervised Learning-Theory and Practice.

This presentation will focus on computational behavioral phenotyping, illustrating how machine learning (ML) methods can be leveraged to transform raw complex multimodal health data into objective and interpretable biomarkers. I first present computer vision and ML algorithms that (i) extract digital behavioral biomarkers from video and gamified visuomotor assessments and (ii) combine them for early autism screening. I then describe computational approaches for analyzing egocentric video to characterize the execution of a chocolate cake cooking task, by producing compact symbolic representations of behavioral trajectories that enable alignment and comparison across patients with sensorimotor and executive disorders. Finally, I present a methodology for learning compact and generalizable representations of wearable biosignals to improve the sensitivity to subtle cardiovascular health changes.

[1] Perochon S., Di Martino J.M. et al. A scalable computational approach to assessing response to name in toddlers with autism. J Child Psychol Psychiatr (2021).

[2] Perochon S., Di Martino J.M. et al. A tablet-based game for the assessment of visual motor skills in autistic children. npj Digit. Med. (2023).

[3] Perochon, S., Di Martino, J.M. et al. Early detection of autism using digital behavioral phenotyping. Nat Med. (2023).

[4] Perochon S., Oudre L. Recursive prototyping for computational behavior analysis from egocentric videos. In review at ICPR 2026

[5] Perochon, S. et al. Time-varying representations of longitudinal biosignals using self-supervised learning. NeurIPS 2024 Workshop: Self-Supervised Learning-Theory and Practice.

¿Cuántas raíces reales tiene un polinomio elegido al azar? En esta charla exploraremos la elegante intersección entre álgebra, probabilidad, estadística y geometría para responder a esta pregunta. Comenzaremos con los resultados históricos de Marc Kac y la motivación de la fórmula de Kac-Rice, para luego adoptar una perspectiva más geométrica mediante la fórmula de Crofton. También daremos el salto al caso de sistemas de ecuaciones polinomiales, motivando el famoso resultado de Shub y Smale. Para finalizar, veremos cómo herramientas como el teorema de tubos de Weyl nos invitan a estudiar las propiedades geométricas de los ceros en sistemas con más variables que ecuaciones. La charla está pensada para ser totalmente autocontenida y dirigida a estudiantes avanzados de diversas licenciaturas, con el objetivo de acercarles las investigaciones actuales que lleva adelante nuestro grupo local.

Edgeworth expansions represent the distribution of normalized sums of i.i.d. random variables as the Gaussian measure corrected by polynomial perturbations, with coefficients expressed in terms of the cumulants of the underlying distribution. 

In this talk, we present an Edgeworth expansion for a class of functionals of a Gaussian field. For an element F of the p-th Wiener chaos, using the Malliavin-Stein method and semigroup analysis, we derive bounds in the total variation distance between the distribution of F and the so-called Edgeworth development of F: a modified Gaussian measure. The bounds depend only on p and the variance of the carré-du-champ operator of F, which governs the proximity of F to the standard normal distribution.