When the topic of Machine Learning arises, it is often accompanied by topics such as artificially intelligent terminator robots taking over the job market. However, the goals in modern Machine Learning differ substantially from making terminator robots. Areas including (but certainly not limited to) computer vision for use in medical imaging and diagnosis have seen tremendous benefit from the rise of Machine Learning in recent years.

Machine Learning refers to the process of making predictions using large amounts of data. For example, given a scan of a human brain, one might be interested in determining a label for the image corresponding to the presence of a certain disease. A machine can learn relationships between images and labels by being trained on a large number of labelled examples.

Machines struggle to learn useful patterns in images by only utilising raw pixel values, as do humans. This type of learning is referred to as being shallow. Humans typically benefit from using features such as regions of discoloration and existence of smaller objects within an image when trying to understand relationships between images and labels. A computer can automatically learn which features of an image are important for predicting its label by using deep learning. Useful features discovered by computers can sometimes be difficult for a human to interpret as the computer interacts with images in purely mathematical ways, whereas humans interact with images simply by looking at them. In recent years Machine Learning engineers have been able to train computers to outperform doctors in medical diagnosis problems; being of great benefit to patient’s whose lives can potentially be turned upside-down by an incorrect diagnosis [1].

Machine Learning complements my degree in mathematics and statistics, making it a natural choice for me to pursue further. Some interesting ideas in Machine Learning and statistics that I had in mind while devising a summer project were:

  • Neural Networks represent functions that can learn sophisticated patterns in data and give accurate predictions on new data.
  • Ensembles combine many Machine Learning algorithms such as Neural Networks to further improve predictive performance on new data.
  • Gaussian Process Regression provides a framework for learning patterns in data while also accounting for uncertainties in making predictions on new data – invaluable for decision making.
  • Bayesian Optimisation is a global optimisation method that utilises Gaussian Process Regression and other ideas from Bayesian statistics.

All of these ideas were combined in my project to create a versatile method that finds the optimal design for an ensemble of Neural Networks to learn sophisticated patterns in data and make accurate predictions, with a view to making implementations on a computer efficient and effective.


[1] Haenssle, H. A., et al, 2018, ‘Man against machine: diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists’, Annals of Oncology, vol. 29, no. 8, pp. 1836–1842.

Jonathan Wilton
University of Queensland

Jonathan Wilton
Contact Us

We're not around right now. But you can send us an email and we'll get back to you, asap.

Not readable? Change text.