School of
Information Technology and Electrical Engineering

Speaker: Dr Alina Bialkowski
Seminar Date: Mon, 05/02/2018 - 11:00
Venue: 49-502; AEB Seminar Room
Host: Prof Amin Abbosh

Seminar Type:  ITEE Research Seminar

Abstract: 

Processing large complex signals often requires the help of machine learning (i.e., algorithms able to learn from and make predictions on data). A lot of the success in machine learning has relied on supervised learning where a huge database of labelled examples is used to learn the mappings from input to output. In real world problems, large sets of ground truth labels are often not available which makes modelling and prediction challenging. In addition, understanding what a model has learnt is often hidden within a black box, making it hard to explain the output predictions, particularly in deep neural network models.

In the first part of the presentation, I will discuss how these challenges can be overcome using my recent work in modelling the perceptual demands of driving as a case study. I will show how this subjective attribute can be quantified though clever ground-truth labelling and video-based deep neural-networks, and then show how to explain and visualise what the model has learnt by inspecting the black-box.

In the second part of the presentation I will demonstrate how these signal processing and machine learning techniques can benefit the biomedical field, and specifically in modelling and imaging the human brain using electromagnetic imaging. Simulation and processing using exact techniques (e.g. integral-based tomography in 3D heterogeneous domains) is computationally expensive with very large memory requirements. Using machine learning, I show how an accurate template of the imaging domain can be generated, to speed up and make the use of established exact algorithms computationally feasible.

Biography: 

Dr Alina Bialkowski is a signal processing and machine learning researcher at The University of Queensland. She completed her BEng in Electrical Engineering (2009) and her PhD (2015) at the Queensland University of Technology. She spent a year at Disney Research Pittsburgh during her PhD, developing algorithms and tools to automatically monitor and analyse team sports, followed by 2.5 years as a postdoctoral researcher at the University College London, using deep neural networks to better understand human perception and attention in driving. Alina has been awarded a best paper prize at WACV 2017 (one of the top applications conferences for machine learning) and four international patents. Alina's research interests are in extracting meaningful information from complex data through feature learning, spatio-temporal data modelling, convolutional neural networks and interactive visualisation.