16 July. 2019, 10:00, Julia House, 21612, CY1591, Themistokli Dervi 3, Nicosia 1066
What do deep neural networks know?
Invited talk by Dr. Olivia Guest, UCL
In this talk I will discuss my work using computational models to understand human and animal cognition as well as how to understand the inner workings of other computational models. How can computational modelling help us open up black boxes and help us understand classification done by pigeons as well as that done by machines? I will explain what computational modelling is and how we can use deep neural networks to model human and animal categorisation. In addition, I will show how these same methods can be used to understand the deep network itself whether we have access to its source code or not analogously to how we can understand classification done by pigeons. I will underline the parallels between understanding black boxes that are incredibly complex feats of human engineering and understanding systems that are products of evolution like brains. Some of the barriers to algorithmic transparency can be overcome using modelling techniques from cognitive science, furthering our understanding of how sophisticated machine learning applications, like deep networks, see the world.
I was born and grew up in Cyprus, but have spent most of the last 15 years in the UK working at the Universities of Oxford, Birkbeck, and UCL. My undergraduate degree was in Computer Science. After that, I moved on to an MSc (2010) and a PhD (2014) focussed on computational modelling for human memory for concepts. I am currently a research scientist in computational modelling at UCL. I create and evaluate computational accounts — such as deep and shallow artificial neural networks — for categorisation and conceptual representation in healthy adults, patient groups, infants, and animal models. I am also interested in using computational modelling and data science broadly in theoretical as well as applied contexts and I am actively for open science and open source. For more, please see: http://oliviaguest.com