Australasian Mathematical Psychology Conference 2019

Modelling human perceptual learning with pre-trained deep convolutional neural networks

Jessica Marris
Melbourne School of Psychological Sciences, University of Melbourne
Amy Perfors
Melbourne School of Psychological Sciences, University of Melbourne
Piers D. L. Howe
Melbourne School of Psychological Sciences, University of Melbourne

With perceptual training, humans can rapidly improve their performance on visual tasks involving complex images (e.g., identifying whether a hip fracture is present in an X-ray image). How do they achieve this? Current models of perceptual learning are typically applied to simple images rather than the complex images that are generally used in real-world tasks. We therefore present a new perceptual learning model based on pre-trained deep convolutional neural networks (DCNNs). Our initial findings show that pre-trained DCNNs can eventually achieve the same level of accuracy as that achieved by humans but do not learn to identify hip fractures in X-ray images as quickly as people do; they require larger training dataset sizes to achieve performance comparable to humans. We hypothesise that one reason for this may be that humans may learn to rapidly identify a hip fracture by drawing on pre-existing concepts that are related to a fracture, such as the concepts of ‘broken’ and ‘unbroken’. The pre-trained DCNNs that we used did not have a visual representation of these concepts as these DCNNs had not been trained to recognise broken images. We therefore investigate whether training DCNNs to differentiate between broken and unbroken objects allows them to learn more quickly to identify hip fractures in X-ray images, thus better simulating human perceptual learning. Our method and preliminary findings will be discussed.