Efficient dynamic neural networks on AR glasses

As machine learning models become larger and more complex, they increasingly demand substan- tial computational resources and memory for train- ing and deployment. Resource-efficient machine learning algorithms address these challenges by op- timizing performance for systems with limited re- sources. With the increasing popularity of wear- able devices such as AR glasses, deploying deep learning models on-device has become essential for enabling real-time and privacy-preserving intelli- gent applications. However, these wearable sys- tems are often constrained in terms of compute, energy, and memory. Resource-Efficient Deep Sub- networks (REDS) provide a promising solution by enabling efficient model adaptation to various dynamic resource constraints without increasing model size. In this project/thesis, you will explore and implement a novel REDS-based visual model deployed directly on the Brilliant AR Glasses. The goal is to demonstrate efficient on-device adapta- tion using REDS, pushing the boundaries of intelli- gent AR applications under dynamic resource con- straints.

Download as PDF

Student Target Groups:

  • Students in ICE, CS or Software Engineering

Thesis Type:

  • Master Project / Bachelor Thesis

Goal and Tasks:

  • Design or select a suitable visual task (e.g., classification, segmentation, gesture recogni- tion) as a demonstration use case;
  • Implement and evaluate REDS-based adap- tation of the selected model under variable real-world conditions;
  • Deploy and benchmark the model on Brilliant AR Glasses using supported toolchains;
  • Present the results and summarize the work in a written report.

Recommended Prior Knowledge:

  • Familiarity with neural networks;
  • Programming skills in Python and C++;
  • Experience with deep learning frameworks (e.g., PyTorch or TensorFlow);
  • (Optional) Knowledge of embedded Android development or interest in mobile hardware acceleration.

Start:

  • a.s.a.p.

Contact: