Danesh Tarapore

This project investigates incremental machine learning for resource-constrained embedded systems, enabling devices to learn from newly collected data without relying on cloud infrastructure or storing complete historical datasets. Building on evolutionary data-subset selection, it develops memory-efficient methods that retain only the most informative examples while minimising catastrophic forgetting as new classes are introduced. The research evaluates the trade-offs between predictive accuracy, memory use, computational cost and energy consumption, with the aim of supporting adaptive, long-lived intelligent devices operating at the extreme edge.