Self Supervised Learning at the Edge: Addressing Label Scarcity on Online Data

Summary:
Edge computing is a new computing paradigm that places powerful computers for analysis, inference, and learning close to physical data sources like sensors, breaking the historical "aggregate your data at a central location and compute" approach.

The standard Artificial Intelligence at the Edge workflow consists of training a model from proven data and then deploying the trained algorithm to make inference at the edge.

Despite this careful preparation, models exhibit two problems when the Edge device collects data:
• the data can drift away from the model and
• the data can be different from the model.

To overcome these challenges, our research has begun to explore Self Supervised Learning, with the goal to eventually allow learning at the edge.

A viable option is to capture valuable information in the synaptic weights of an artificial neural network through a technique called Self Supervised Federated Learning. Our research team is applying new Self-supervised techniques to two different datasets collected by Sage nodes - passively recorded soundscapes and full-sky images.

This talk will show how we can autonomously extract relevant features from images of the sky and from audio spectrograms without human produced labels by
• Separating clouds and clustering them based on features, and
• Separating silence and background noise from sounds of interesting events.

This novel approach requires reduced human intervention, suggesting a new path for data characterization at the edge.

About the Speaker:
Dario Dematties is Postdoctoral Researcher at Northwestern-Argonne Institute of Science and Engineering.
Dario's research focuses on Self-supervised Learning (SSL) and Federated Learning (FL) for Edge Computing (EC) scenarios. He is applying SSL in the context of automatic atmospheric conditions characterization and in avian diversity monitoring. He has also collaborated with Uppsala University researchers in Sweden by applying Deep Learning to Nanopore Translocation Signal Processing.

Dario received his Electronic Engineering degree from Universidad Tecnológica Nacional (UTN) in Mendoza, Argentina in 2012 and his PhD degree at the Instituto de Ingeniería Biomédica, Facultad de Ingeniería, Universidad de Buenos Aires, Argentina in 2020. His PhD thesis focused on Biological Plausible modeling of early human language acquisition. Later at the National Scientific and Technical Research Council (CONICET) Argentina, he worked as a Postdoctoral Researcher on Bio-inspired Foveated Computer Vision using Deep Learning.

Agenda:
(Times are Central Daylight Time)
6:00pm - brief intros
6:05pm - Talk by Dario Dematties
6:50 pm – Q&A
7:00 pm - end

Registration for Zoom
Note: You must register with Zoom for the remote broadcast, registering in Meetup does not give you access to the Zoom meeting.