Understanding Performance Trade-Offs in Low Power Artificial Intelligence

AI Quick Summary

One solution to limit the bandwidth requirements for intelligent IoT systems is to push intelligence to the "edge". The trade-off here is that processing on end-devices increases their cost and computational power requirements. This can be particularly disadvantageous for battery powered devices which have no connected power source, driving a new a need for low power AI. This project examines the trade-offs between performance vs computational power in such low power AI systems. Initial situation and problem definitionIt is estimated that over 30B devices were connected to the internet by 2020 [1]. This number is growing fast as the so-called Internet-of-Things expands. Many first generation IoT solutions send data back to the cloud for processing. This requires reliable high bandwidth communications networks to support all of this traffic. One solution to limit the growing bandwidth requirements for intelligent IoT systems is to push "intelligence" i.e. the computational load, to the "edge". The trade-off here is that processing on end-devices increases their cost and computational power requirements. This can be particularly disadvantageous for battery powered devices which have no connected power source, driving a new a need for low power AI [2]. This project examines the trade-offs between performance vs computational power in such low power AI systems.Aim of the Work and Expected ResultsThere are two main goals of this research work. The first is based on literature and desk research (primary and/or secondary). In this part, the student is expected to understand the current state of the art in low power AI for IoT systems, and in particular for machine learning models. The second goal of the project is based on actual implementation and exploration of machine learning architectures to measure the performance/power trade-off in a specific setting. As a result of these two pieces of work, it is expected that the student will be able to present useful findings and recommendations on low-power AI systems, and most desirably, performance charts which allow system designers to pick the optimal trade-off for their architecture between computational power and system performance.Desired Methods & ProcedureFor the state-of-the art research, the student should engage in a classic literature research using e.g. Google Scholar and the resources form the HSLU library (swisscovery.ch). The student is epected to read and write articulately in the English language. The litereature review shoudl be up to date and at a high technical level. It is desired that students have taken the module STW or WSFM on Scientific and Technical Writing previous to or in parallel with the Bachelor thesis. For the machine learning model research the students are expected to work in the Python programming language using standard libraries such as Scikit-Learn, Pytorch, Keras, or Tensorflow. For development the student will work most likely with Jupyter Notebooks or the favourite IDE such as Spyder or Pytorch etc. Candidate Students are expected to have taken and passed the Module ML or ADML (Advanced) Machine Learning at HSLU.Creativity, Variations & InnovationThe scope for creativity in this module lies in potentially finding new architectures or model pruning methods which can advance the state of the art in low power AI. This is a relatively new and growing field with huge potential scope for application in industry. Therefore the student is encouraged to be innovative and try new ways to squeeze more from less in a low power IoT environment.References[1] Zou, Zhuo, et al. "Edge and fog computing enabled AI for IoT-an overview." 2019 IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS). IEEE, 2019.[2] Ye, Le, et al. "The challenges and emerging technologies for low-power artificial intelligence IoT systems." IEEE Transactions on Circuits and Systems I: Regular Papers (2021).

Project Idea Metadata

Project Idea Name: Understanding Performance Trade-Offs in Low Power Artificial Intelligence
Date: 11/8/2021 12:12:41 PM
Administrators:
- Donnacha Daly

Project Idea Description

Initial situation and problem definition

It is estimated that over 30B devices were connected to the internet by 2020 [1]. This number is growing fast as the so-called Internet-of-Things expands. Many first generation IoT solutions send data back to the cloud for processing. This requires reliable high bandwidth communications networks to support all of this traffic. One solution to limit the growing bandwidth requirements for intelligent IoT systems is to push "intelligence" i.e. the computational load, to the "edge". The trade-off here is that processing on end-devices increases their cost and computational power requirements. This can be particularly disadvantageous for battery powered devices which have no connected power source, driving a new a need for low power AI [2]. This project examines the trade-offs between performance vs computational power in such low power AI systems.

Aim of the Work and Expected Results

There are two main goals of this research work. The first is based on literature and desk research (primary and/or secondary). In this part, the student is expected to understand the current state of the art in low power AI for IoT systems, and in particular for machine learning models. The second goal of the project is based on actual implementation and exploration of machine learning architectures to measure the performance/power trade-off in a specific setting. As a result of these two pieces of work, it is expected that the student will be able to present useful findings and recommendations on low-power AI systems, and most desirably, performance charts which allow system designers to pick the optimal trade-off for their architecture between computational power and system performance.

Desired Methods & Procedure

For the state-of-the art research, the student should engage in a classic literature research using e.g. Google Scholar and the resources form the HSLU library (swisscovery.ch). The student is epected to read and write articulately in the English language. The litereature review shoudl be up to date and at a high technical level. It is desired that students have taken the module STW or WSFM on Scientific and Technical Writing previous to or in parallel with the Bachelor thesis. For the machine learning model research the students are expected to work in the Python programming language using standard libraries such as Scikit-Learn, Pytorch, Keras, or Tensorflow. For development the student will work most likely with Jupyter Notebooks or the favourite IDE such as Spyder or Pytorch etc. Candidate Students are expected to have taken and passed the Module ML or ADML (Advanced) Machine Learning at HSLU.

Creativity, Variations & Innovation

The scope for creativity in this module lies in potentially finding new architectures or model pruning methods which can advance the state of the art in low power AI. This is a relatively new and growing field with huge potential scope for application in industry. Therefore the student is encouraged to be innovative and try new ways to squeeze more from less in a low power IoT environment.

References

[1] Zou, Zhuo, et al. "Edge and fog computing enabled AI for IoT-an overview." 2019 IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS). IEEE, 2019.

[2] Ye, Le, et al. "The challenges and emerging technologies for low-power artificial intelligence IoT systems." IEEE Transactions on Circuits and Systems I: Regular Papers (2021).