By Brent Martin, adjunct senior research fellow, University of Canterbury
Machine learning (a branch of Artificial Intelligence) is the general term for how computers learn to recognise patterns. Machine Learning has become very powerful with the advent of inexpensive computer processing and the explosion in available data enabled by the internet and cloud computing. Capitalising on all this data, deep neural networks have recently shown order-of-magnitude improvements in tasks such as automatic speech recognition and computer vision.
Machine learning is potentially very useful for the cacophony project for the following tasks
- Predator identification from video
- Learning lures
- Sound identification of predators
There is a real issue with current methods of measuring pest numbers as they often rely on tracking tunnels and chew cards which many animals currently avoid. Video has proven to be the most robust way to be sure to know which predators are out there but it is a very manual process to got through each video and identify the animals. The aim of our project is to use machine learning to do this.
Initially the machine learning will work only on the cloud but eventually it could be run on a device like a smart phone. Training the models will always be calculated in the cloud as this requires a large amount of computing power but applying the model is possible on smaller devices such as smart phones.
The aim of the project is a tool that allows anyone to upload video and have it automatically classify into one of the following groups
Later it may be able to also extent to monitor mice, hedgehogs, and possibly different sexes as well.
There are going to be a large range of different sounds and images that are going to be tested and it may be possible to create a system that merges two different lures to create a new type that could be better than the previous one. This is a project for later in the process…
Sound identification of predators
It is possible to have a continual loop of sound recording that only saves before each animal sighting. This could be analysed to learn what sounds are the most likely to indicate a predator is in the area. This learning algorithm could then be running in the background to work out what is the best lure to play. This could be combined with time, season, weather, habitat etc to learn when best lures to try. This is another project that we will try later in the project.
The first project to identify pests is being run by Brent Martin and Kourosh Neshatian From the Computer Science and Software Engineering department of the University of Canterbury. A group of final year honour students will apply deep neural networks to the problem of video classification, with the hope of developing learning algorithms that can reliably classify predators in images. All of the work will be open source and designed to be built upon with future iterations. First versions of the project should be completed in June 2016.