EDGE DEPLOYMENT AND CONTINUAL UPDATING OF FOUNDATION MODELS

Introduction

Deploying large pretrained models on edge devices is attractive but difficult because deployment and adaptation must happen under strict compute, energy, and memory constraints. Recent on-device learning literature shows that practical edge adaptation requires careful simplification, selective updates, and efficient memory handling. Replay-based continual learning and lightweight update schemes are particularly relevant here.

Objectives

·  Deploy a compact foundation model, LLM/VLM component, or sensor model on an edge-class device or emulator.

·  Investigate lightweight continual update strategies.

·  Quantify the trade-offs between adaptation quality and system cost.

Tasks

1.     Literature Review: Edge AI, parameter-efficient adaptation, continual learning on-edge device.

2.     Model Selection: Choose a realistic edge-compatible foundation model or compressed pretrained model.

3.     Deployment: Run the model on an embedded board, mobile platform, or realistic edge simulator.

4.     Continual Updates: Compare state-of-the-art low-cost adaptation methods such as methods based on replay, adapters, LoRA-style tuning, or sparse updates.

5.     Evaluation: Measure task performance, latency, memory, energy, and forgetting.

6.     Optional Extension: Add drift-triggered adaptation rather than continuous updating.

Pre-requisites

Python, deep learning, systems/programming interest, embedded/edge experience helpful.

Work

15% Theory, 65% Programming/Experiments, 20% Writing

Contact

Ali Sabzi Khoshraftar (a.sabzikhoshraftar@utwente.nl)