Project Summary
Humans are expert categorisers. We can sort any number of stimuli into myriad categories, often after minimal learning. For example, we can easily judge a piece of fruit as either an apple or a pear, but we could also categorise the same object in many other ways: Is it red or green? Ripe or unripe?Recent studies on the brain basis of categorisation have come largely from invasive recordings in animals. Understanding the neural mechanisms of how this happens in humans is comparatively less well understood. One way to make progress on this issue it to train artificial neural networks to perform categorisation tasks – this allows us to look ‘under the hood’ at how the artificial system solves the categorisation problem and to develop hypotheses about the process in human brains.The purpose of this internship will be to train artificial neural networks to become expert categorisers and compare their behaviour to that of the human brain. You will learn about various theories of how mammalian brains learn to categorise efficiently, gain experience in using modern machine learning tools to train artificial neural networks, and learn how to examine neural data recorded from humans while they categorise visual stimuli.
Full Project Details
Humans show an extraordinary ability to categorise flexibly any number of stimuli. For example, we can easily judge a piece of fruit as either an apple or aa pear, but we could also categorise the same object in many other ways: Is it red or green? Ripe or unripe?
Neurophysiological data suggest that this ability may rely on parts of the brain that are dedicated to categorising. These neural populations may multitask – that is, they are used for making all sorts of judgments, from telling cats apart from dogs to judging whether a piece of fruit is ripe or not (Cromer et al., 2010; Summerfield et al., 2020). However, this neural efficiency may come at a cost to flexibility (Roy et al., 2010; Chaisangmongkon et al., 2017). This project aims to use brain imaging and machine learning to investigate this apparent trade-off. You will compare behavioural measures and neural data recorded from human participants engaged in flexible categorisation with artificial neural networks trained to perform the same categorisation task. The aim is to learn more about the trade-off between generality and flexibility and whether it is influenced by the learning context.
You will have the opportunity to simulate the tasks typically performed by humans and animals in recurrent neural nets and relate these simulations to existing empirical data. You will be able to learn how to perform advanced pattern analysis of neural data (Hall-McMaster et al., 2019; Myers et al., 2015) or fit computational models of category-learning to behaviour, and relate these to the artificial neural models of categorisation.
Further Reading:
Chaisangmongkon, W., Swaminathan, S. K., Freedman, D. J., & Wang, X. J. (2017). Computing by robust transience: how the fronto-parietal network performs sequential, category-based decisions. Neuron, 93(6), 1504-1517.Cromer, J. A., Roy, J. E., & Miller, E. K. (2010). Representation of multiple, independent categories in the primate prefrontal cortex. Neuron, 66(5), 796-807.Hall-McMaster, S., Muhle-Karbe, P. S., Myers, N. E., & Stokes, M. G. (2019). Reward boosts neural coding of task rules to optimize cognitive flexibility. Journal of Neuroscience, 39(43), 8549-8561.Myers, N. E., Rohenkohl, G., Wyart, V., Woolrich, M. W., Nobre, A. C., & Stokes, M. G. (2015). Testing sensory evidence against mnemonic templates. Elife, 4, e09000.Roy, J. E., Riesenhuber, M., Poggio, T., & Miller, E. K. (2010). Prefrontal cortex activity during flexible categorization. Journal of Neuroscience, 30(25), 8519-8528.Summerfield, C., Luyckx, F., & Sheahan, H. (2020). Structure learning and the posterior parietal cortex. Progress in Neurobiology, 184, 101717.