Capstone Hall of Fame

Implementation of Knowledge Transfer in Machine Learning

Effectively transferring previously learned knowledge to a new domain is one of the hallmarks of human intelligence (e.g., people good at playing the guitar are more likely to learn ukulele faster). As an analogy in machine learning, transfer learning is a rising technique where we develop a model for a new task through the knowledge learned from a related but different task. Generally, abundant data is needed to train a model for a task but access to that data is not always available. Transfer learning is particularly useful when there is insufficient data for a new problem and there is a big pre-existing data pool that can be transferred to the problem. As a consequence, transfer learning can be vital in certain applications such as digit recognition, text classification, and medical imaging where obtaining large data sets for training models is either costly or impossible.

In this project, you will become familiar with general transfer learning methodologies, and then develop novel and efficient transfer learning algorithms for general machine learning problems, such as image recognition. For example, in training a classifier to predict whether an image contains food, you could use the knowledge it gained during training to recognize drinks. You can use statistical learning methods or construct deep neural networks to achieve effective knowledge transfer.

MUR Accumulator

The Accumulator team running in semester 2 2021 aims to make a head start on the design and implementation of the 2022 electric race car for the Melbourne University Racing team in collaboration with the mechanical chassis team. The project will aim to improve efficiencies and reliability in the design of a new battery pack whilst allowing for early completion of the 2022 vehicle to extend the testing timeframe. With the aid of the HV team already looking at the feasibility of regenerative braking on a Formula SAE electric race car, the team will aim to reduce the energy capacity and hence pack size whilst utilising energy efficiency measures. They will explore optimisation of battery management and cell balancing, as well as facilitate the introduction of active cooling to ensure safe operation of the vehicle. With the MUR team looking to optimise their performance at the competition, all designs will need to comply with the FSAE-Australasia rules published annually, whilst innovations explored by the team will be encouraged for the progression of the automotive industry.

Bluetooth Hearing Device

Attention problems are common in the classroom. Many children have trouble clearly hearing the teacher’s voice above the background classroom noise. Commonly used strategies, including stimulant medication have side effects. An alternative approach is to use a remote microphone assistive listening device (FM radio/hearing aide), which boosts the signal-to-noise ratio for the listener (child). Currently available devices are expensive and can attract a stigma when worn in the classroom.

Students are invited to apply for this collaborative project to create a Bluetooth-based hearing assistance device. The challenge is to create a working protype with a mobile app that syncs a Bluetooth microphone (worn by the teacher) with Bluetooth headphones (worn by the child). Students will be asked to optimise audio latency and syncing.

Demand Analysis & Solar Energy Forecasting Using Real-time Data

As sustainability concerns and climate change become a part of everyday conversation, it is a rewarding time to get involved in improving how solar PV panels can be more efficient and making them a more desirable asset. With this project, the students will have an exciting opportunity to work with real-time data collected from SolShare, the world’s only hardware for sharing rooftop solar to multi-tenanted apartments. SolShare actively monitors demand and solar PV generation, and optimally distributes solar energy from a single rooftop system to multiple units. In this project, the students will work on demand side data and solar energy data. They will learn and use signal processing and machine learning algorithms to identify which house appliances are being used at any time to drive demand response management and forecasting techniques for short-term solar forecasting to improve the energy distribution algorithm. Then, they will develop optimisation algorithms to dynamically share solar energy to minimise export to the grid and maximise customer utilisation. These algorithms will then be tested in a simulation bench set-up to confirm their performance. After successful testing, the algorithms will be potentially used in actual products operating all over Australia, Europe, and the USA.