Mixed Selectivity via Unsupervised Learning in Neural Networks

Mixed selectivity characterises neurons that simultaneously respond to different input stimuli. Neurons with mixed selectivity have been observed in multiple brain regions, and are hypothesised to play important roles in neural computation. Recently, both experimental and theoretical work demonstrated the importance of mixed selectivity in context-dependent decision tasks. This thesis extends existing theoretical work on mixed selectivity, arguing for a general and statistical role of mixed selectivity in learning complex dependencies of input stimuli. This role can be motivated from unsupervised learning of generative models, and is exhibited in increased mutual information between the stimuli and their neural representation. This argument is supported empirically using simulation of a sequence disambiguation task that incorporated key aspects of related behaviour experiments. Mixed selectivity neurons that resembled hippocampal place cells were emerged from models optimised only for behaviour. To understand these results, as well as to generalise the findings to a wider range of computations, I provided a formal connection between learning robust models and mixed selectivity.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2019

Author dc:creator

• Wu, Yan

Advisor dc:contributor.advisor

• Lengyel, Mate

• Neural Network
• Unsupervised Learning
• Neural Representation
• Information Theory
• Machine Learning

Licence

• https://www.repository.cam.ac.uk/bitstreams/0153a149-9917-478a-98c8-d26aa0a6481a/download
• https://www.rioxx.net/licenses/all-rights-reserved/

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.53941

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/306850