Research
Projects
Federated and continual learning from heterogenous data in devices and robots

Federated and continual learning from heterogenous data in devices and robots

Fast adaptation, personalization, and data privacy are three important aspects highly valued and required in most devices and robots. With billions of mobile devices in the world and sensors everywhere, this is an era of ubiquitous computing, with an inconceivable and overwhelming explosion in the volumes of data that are being generated. In this context, federated learning will have a huge impact, as it will favor an even deeper penetration of machine learning in all types of devices and robots, and in a broader spectrum of applications. Federated learning is a strategy for training and sharing models amongst machines in general (devices, robots), but without the exchange of data. The basic federated learning process consists of two main steps: (1) local adaptation of models on the devices or robots, and (2) global consensus of the models in the cloud or central server. Hence, there is a cyclical process of local adaptation (in the devices), and global consensus (in the cloud). What is transferred to the cloud are models from which data cannot be retrieved. This satisfies privacy concerns, requires less bandwidth of communications, but still keeps the benefits of sharing knowledge. Federated learning offers an extremely powerful and novel possibility that is called personalized federated learning. Briefly, this means that devices that are in a federation and share the models, will learn something that is halfway between the global model reached by consensus, and the models that they would had learned locally only from the data captured on the devices themselves. Two important challenges that we will face in this project are related to the creation of these personalized federated models, in particular, we will design algorithms so achieve personalized federated learning that (1) is robust to data heterogeneity amongst the different devices (non i.i.d. data), and (2) is able to achieve continuous federated learning (i.e., deal with nonstationary data or different tasks).

We are completely certain that federated learning will revolutionize learning in all machines, robots included. In the case of robotics, we believe that the use of federated learning has an enormous potential as well. Due to this, we will tackle with a third challenge: how to extend deep reinforcement learning to the context of a personalized federated scenario. This third challenge is particularly suitable for robots, and it has been not enough explored yet. To perform the desired research, we will solve three important tasks (technological demonstrators): "Learning contact-based industrial tasks through dexterous robotic manipulation", "Continual object learning from low-labelled stream video sequences" and, finally, "Human activity recognition on smartphones". The first demonstrator will allow us to collaborate with the Engineering Graduate School SIGMA-Clermont and the Institut Pascal Laboratory (Clermont-Ferrand, France), and even collaborate in two European H2020 Projects: ACROBA (which will begin in January 2021) and SoftManBot (that began in October 2019). The second technological demonstrator has got the attention of a company so important as Televés. And finally, the third demonstrator is important for Situm Technologies, an important spin-off of the University of Santiago de Compostela.

Objectives

Objective 1: Development of strategies to enable federated learning with heterogeneous data

Objective 2: Development of strategies to facilitate continuous federated learning.

Objective 3: federated learning applied to deep reinforcement learning, and to the specific context of robotics

Objective 4: ncremental Learning from Low-labelled Stream Data in Open-world recognition