Description:

Machine learning models are increasingly becoming larger as we try to capture complex behaviour. However, with this increasing complexity it can become difficult, or even impossible, to run such networks on typical embedded hardware. Instead, what methods can we apply to “optimise” these networks to make them simpler and more efficient?

This project will investigate approaches to simplify the structure of pre-trained MLP models to reduce their computational complexity while maintaining high accuracy. This may be through the use of a range of approaches such as pruning (removing the least-relevant components) or subdividing (splitting the network up into several smaller networks).

Through this approach, engineers would be able to train complex models offline on powerful hardware, and then create simplified versions that can be distributed for execution.

Type:

Undergraduate

Outcome:

Automatic process to reduce the complexity of pre-trained MLPs, with two (potential) objective functions:

1. Ability to reduce an MLP to a fixed size to ensure real-time execution
2. Ability to reduce an MLP to a minimal size that ensures high accuracy

Prerequisites

None

Specialisations

• Computer Systems Engineering
• Electrical and Electronic Engineering
• Software Engineering

Categories

• AI & Machine learning
• Embedded Systems

Supervisor

• Nathan Allen

Co-supervisor

• Partha Roop

Team

• James Beazley
• Penelope Goertzen

Lab

Embedded Systems (405.760, Lab)