Project Description
This project aims to bridge the gap between MATLAB-based machine learning (ML) development and real-world deployment on embedded edge devices. Matlab provides powerful and user-friendly machine-learning packages for the fast development of AI model training, facilitating edge AI research in terms of model generality and scalability. The current obstacle is that most embedded platforms of edge AI, such as Jetson Nano, do not support the Matlab language, so an efficient pipeline to convert AI algorithms from Matlab to Nano-compatible language is essential. This project aims to build such a pipeline to bridge the gap between MATLAB-based machine learning (ML) development and real-world deployment on embedded edge devices. Key objectives include optimizing and deploying ML models developed in MATLAB for efficient execution on the NVIDIA Jetson Nano. A streamlined workflow should be established for the language conversion of all ML algorithms developed. To achieve these objectives, this project will implement ML models in MATLAB to leverage its built-in tools (e.g., Classification Learner, Deep Learning Toolbox), convert MATLAB-trained models to formats compatible with Jetson Nano (e.g., ONNX, TensorRT), generate compatible algorithms using MATLAB’s GPU Coder, and deploying the models on Jetson Nano for real-time inference.
Tasks and Responsibilites
The student involved in this project will gain hands-on experience in machine learning deployment on embedded systems and contribute to key research activities, including
(1) Implement and train machine learning models using MATLAB’s Deep Learning Toolbox and Statistics and Machine Learning Toolbox.
(2) Convert trained MATLAB algorithms for compatibility with Jetson Nano by using MATLAB GPU Coder and deep learning tools to optimize models for embedded deployment.
(3) Implement and test real-time inference on Jetson Nano for performance evaluation by measuring deployed models' inference speed, memory usage, and power consumption.
(4) Experiment with hardware acceleration using CUDA and NVIDIA Deep Learning SDK to enhance edge AI performance.
(5) Maintain a detailed log of experimental results, model configurations, and optimization techniques. Prepare progress reports and presentations summarizing findings.