Epifanios Baikas

MINDS CDT doctoral researcher | Senior Research Assistant, University of Southampton

Making AI learn directly on the devices where data is created

Formal project title: Efficient learning on resource-constrained embedded systems

Epifanios Baikas is developing machine-learning methods that allow Internet of Things devices to train and update AI models locally. By reducing the need to send sensitive data to cloud servers, his research aims to strengthen privacy, reduce dependence on network connectivity and help intelligent devices adapt to individual users.

Profiles: LinkedIn | Google Scholar | ORCID

Research at a glance

Research focus	On-device machine learning
MINDS theme	Embedded Artificial Intelligence
Technologies	Machine learning, Internet of Things, edge devices
Wider applications	Privacy-preserving and personalised AI
Current status	Doctoral researcher
Related role	Senior Research Assistant, University of Southampton

Training AI without sending data to the cloud

Internet of Things devices collect growing amounts of information about the people and environments around them. AI models commonly process and learn from this information using cloud infrastructure, which means data must leave the device and travel across a network.

Epifanios Baikas is investigating an alternative: enabling an AI model to train, update and personalise itself directly on the device where the data is created.

This is technically challenging because many Internet of Things devices have extremely limited memory and computing capacity. Conventional machine-learning methods assume that a complete dataset is available and can be divided into balanced batches. On an edge device, information may instead arrive incrementally, with data from the same class grouped together and the total number of classes initially unknown.

His research develops methods that work within these constraints, allowing models to learn from data arriving over time on memory-restricted hardware.

Why on-device learning matters

Research approach

The research examines how machine-learning models can train using small batches of data arriving incrementally on devices with restricted volatile and non-volatile memory. It focuses particularly on data that do not follow the independent and identically distributed assumptions used by conventional training methods.

Research keywords: On-device learning, edge AI, Internet of Things, memory-constrained computing, incremental learning, privacy-preserving AI

Research and outputs

Featured output: GECCO 2024 conference paper

Parts of Epifanios’s doctoral research were presented at GECCO 2024, the Genetic and Evolutionary Computation Conference.

View the GECCO 2024 conference paper in the ACM Digital Library

Research profiles

• Epifanios Baikas on Google Scholar
• Epifanios Baikas on ORCID
• Epifanios Baikas on LinkedIn

Extending the research into neuromorphic hardware

Alongside completing his PhD, Epifanios Baikas is a Senior Research Assistant in the University of Southampton’s Cyber-Physical Systems research group.

His work on the Edgy Organism project builds on the same underlying challenge as his doctoral research: helping AI learn efficiently on constrained hardware. The role develops this further through work on neuromorphic hardware systems, on-device learning and spiking neural networks.