Epifanios Baikas MINDS CDT, 2022

Doctoral Researcher and Senior Research Assistant, University of Southampton

Epifanios Baikas is a doctoral researcher working on machine learning for Internet of Things devices, with a focus on a practical privacy problem that sits at the centre of how those devices operate.

Most IoT devices collect data continuously and send it to cloud servers, where AI models are trained and updated. That process raises real concerns: data leaves the device, passes through networks, and is processed remotely. Epifanios's research is developing methods that allow AI models to learn and update directly on the device itself, eliminating the need to transmit any sensitive data to the cloud.

The technical challenge is significant. IoT devices typically have very limited memory, often measured in hundreds of kilobytes, which makes storing and processing training data on the device genuinely difficult. Conventional training methods assume that a complete dataset is available and can be split in balanced, varied batches. However, on an IoT device data tend to arrive in sequences of uniform batches, which violates standard statistical assumptions made by conventional training methods and requires different approaches entirely.

His research investigates how to work with batches of data arriving incrementally on memory-restricted hardware, developing methods that allow models to train, update and personalise locally even without a network connection. That last point matters: a model that does not need the cloud to function can serve users in environments where connectivity is limited or unavailable.

Parts of his work were presented at GECCO 2024, a leading conference in genetic and evolutionary computation.

Alongside his doctoral research, Epifanios has taken up a role as Senior Research Assistant within the Cyber-Physical Systems research group at the University of Southampton, working on the Edgy Organism project. This work builds directly on his PhD research, with a greater focus on neuromorphic hardware systems and on-device learning approaches for spiking neural networks.