IEM Seminar Series: Machine Learning for plant identification and other challenges from citizen science
Event details
| Date | 03.04.2023 |
| Hour | 10:00 › 11:00 |
| Speaker |
Prof Joseph Salmon Université de Montpellier, France |
| Location | Online |
| Category | Conferences - Seminars |
| Event Language | English |
Abstract
In supervised learning --- for example for image classification --- modern massive datasets are usually labeled by a crowd of "workers".
An example of interest is the case of citizen science, where data is potentially provided and annotated by volunteers. In particular, we will describe the learning challenges related to an application such as Pl@ntnet, aiming at plant species recognition with a simple cell phone leveraging both expert and amateur botanists.
The labels obtained in such a framework must be aggregated to improve learning.
This aggregation step is usually based on a confidence score per annotator.
However, these annotator-centric approaches do not take into account the ambiguity of each task.
Some inherently ambiguous tasks may even mislead expert workers, which could potentially hinder the learning step.
We will present an adaptation - to participatory learning scenarios - of the AUM (Area Under the Margin), a tool for identifying ambiguous tasks, which allows for improved generalization or calibration performance.
Bio
Joseph Salmon is a full professor at Univ. Montpellier since 2018.
Prior to joining Univ. Montpellier, he was a visiting professor at UW from February to August 2018, and an assistant professor at Télécom ParisTech from 2012 to 2018.
His research interests include cooperative learning, convex optimization, sparse regression models and inverse problems from imaging science.
His recent contributions have aimed at improving machine learning techniques for plant identification, leveraging crowdsourcing and citizen science.
Prior works have focused on image denoising (in particular for low photon counts), on speeding-up standard Lasso solvers and on handling the noise structure to improve signal estimation.
As an associate member of the INRIA Parietal Team, he was contributing to applying his work to brain imaging challenges.
In supervised learning --- for example for image classification --- modern massive datasets are usually labeled by a crowd of "workers".
An example of interest is the case of citizen science, where data is potentially provided and annotated by volunteers. In particular, we will describe the learning challenges related to an application such as Pl@ntnet, aiming at plant species recognition with a simple cell phone leveraging both expert and amateur botanists.
The labels obtained in such a framework must be aggregated to improve learning.
This aggregation step is usually based on a confidence score per annotator.
However, these annotator-centric approaches do not take into account the ambiguity of each task.
Some inherently ambiguous tasks may even mislead expert workers, which could potentially hinder the learning step.
We will present an adaptation - to participatory learning scenarios - of the AUM (Area Under the Margin), a tool for identifying ambiguous tasks, which allows for improved generalization or calibration performance.
Bio
Joseph Salmon is a full professor at Univ. Montpellier since 2018.
Prior to joining Univ. Montpellier, he was a visiting professor at UW from February to August 2018, and an assistant professor at Télécom ParisTech from 2012 to 2018.
His research interests include cooperative learning, convex optimization, sparse regression models and inverse problems from imaging science.
His recent contributions have aimed at improving machine learning techniques for plant identification, leveraging crowdsourcing and citizen science.
Prior works have focused on image denoising (in particular for low photon counts), on speeding-up standard Lasso solvers and on handling the noise structure to improve signal estimation.
As an associate member of the INRIA Parietal Team, he was contributing to applying his work to brain imaging challenges.
Practical information
- General public
- Free