View Proposal

---

Proposer

Patricia Vargas

Title

Federated Learning for Social Robots performing Activity Recognition in Ambient Assisted Living environments.

Goal

To use Federated Learning for Social Robots to perform Activity Recognition in Ambient Assisted Living environments

Description

In the context of assistive technologies for the elderly or people with disabilities, intelligent environments have been designed to empower this public with more autonomy in daily activities. To this aim, sensors and actuators embedded in the environment or in social robots might be orchestrated to produce helpful behaviours. In any case, deploying this type of technology requires that the system identifies the state of the environment and the users. This understanding can be achieved through activity recognition methods, many of which are presented in the literature with good results in several applications. However, these methods usually require data from several users to be concentrated in a centralised computational base to train machine learning models. This requirement, especially considering modalities such as videos or audio, raises ethical and legal concerns regarding data privacy. In this work, we propose to train the models locally for each participant using a Federated Learning approach to induce models based on public datasets such as the HWU-USP dataset. This approach preserves privacy because it does not require that the data is transferred out of the user’s environment, only partially trained models. Metrics such as time elapsed and accuracy will be evaluated.

Resources

Python, Machine learning basics, at least one machine learning framework.

Background

Deep learning, Federated learning, Activity recognition, Ambient assisted living

Url

Difficulty Level

Challenging

Ethical Approval

Full

Number Of Students

1

Supervisor

Patricia Vargas

Keywords

python, machine learning basics, at least one machine learning framework

Degrees

Master of Science in Robotics