Neural mechanisms for flexible behaviour in humans and artificial neural networks
The natural environment is non-stationary – whilst our daily life often follows a routine, circumstances can change, and no two experiences are the same. To thrive, intelligent agents must therefore be able to flexibly adjust the way they process information to support the pursuit of their goals. In...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2023
|
| Subjects: |
| _version_ | 1826313338999537664 |
|---|---|
| author | Piwek, EP |
| author2 | Summerfield, C |
| author_facet | Summerfield, C Piwek, EP |
| author_sort | Piwek, EP |
| collection | OXFORD |
| description | The natural environment is non-stationary – whilst our daily life often follows a routine, circumstances can change, and no two experiences are the same. To thrive, intelligent agents must therefore be able to flexibly adjust the way they process information to support the pursuit of their goals. In this thesis, I explored different aspects of such flexible, intelligent behaviour. To this end, I used a combination of artificial neural network modelling and human behavioural methods, comparing network activity patterns and behaviour against their biological counterparts. In the first set of experiments, I examined how artificial agents manipulate information in their short-term memory to match changing task demands. I demonstrated that neural networks used two distinct representational formats to prevent cross-interference and support the generalisation of information across contexts. Crucially, these representations were a striking match to those previously observed in biological brains. I then built on this result by analysing aspects of the neural network model not readily accessible to biological researchers, revealing the mechanistic and normative reasons for the observed results. In the second experiment, I explored how information can be prioritised in short term memory in situations where it can prove to be irrelevant for guiding behaviour at a later stage. I showed that artificial agents achieved this goal by amplifying the proportion of memory resources dedicated to prioritised information maintenance, without transforming it into a format that could guide subsequent behaviour. In the third experiment, I examined how logical categories can be learnt from an algorithmic perspective. I showed that a behavioural marker previously thought to index rule-based processes involving symbolic representations in humans was equally predictive of the behaviour of artificial agents not equipped with such features. In the last experiment, I investigated how humans learn simple problems and showed that they might have an innate predisposition to search for abstract rules, allowing them to generalise their knowledge to new contexts. Taken together, it is hoped that the work presented in this thesis furthers our understanding of the principles behind flexible, intelligent behaviour and sheds new light on their mechanistic implementation in neural circuits. |
| first_indexed | 2024-09-25T04:11:28Z |
| format | Thesis |
| id | oxford-uuid:01cfc86b-3446-4e8d-9d11-9b970024e408 |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-09-25T04:11:28Z |
| publishDate | 2023 |
| record_format | dspace |
| spelling | oxford-uuid:01cfc86b-3446-4e8d-9d11-9b970024e4082024-07-01T11:55:04ZNeural mechanisms for flexible behaviour in humans and artificial neural networksThesishttp://purl.org/coar/resource_type/c_db06uuid:01cfc86b-3446-4e8d-9d11-9b970024e408Cognitive neuroscienceComputational neuroscienceEnglishHyrax Deposit2023Piwek, EPSummerfield, CStokes, MThe natural environment is non-stationary – whilst our daily life often follows a routine, circumstances can change, and no two experiences are the same. To thrive, intelligent agents must therefore be able to flexibly adjust the way they process information to support the pursuit of their goals. In this thesis, I explored different aspects of such flexible, intelligent behaviour. To this end, I used a combination of artificial neural network modelling and human behavioural methods, comparing network activity patterns and behaviour against their biological counterparts. In the first set of experiments, I examined how artificial agents manipulate information in their short-term memory to match changing task demands. I demonstrated that neural networks used two distinct representational formats to prevent cross-interference and support the generalisation of information across contexts. Crucially, these representations were a striking match to those previously observed in biological brains. I then built on this result by analysing aspects of the neural network model not readily accessible to biological researchers, revealing the mechanistic and normative reasons for the observed results. In the second experiment, I explored how information can be prioritised in short term memory in situations where it can prove to be irrelevant for guiding behaviour at a later stage. I showed that artificial agents achieved this goal by amplifying the proportion of memory resources dedicated to prioritised information maintenance, without transforming it into a format that could guide subsequent behaviour. In the third experiment, I examined how logical categories can be learnt from an algorithmic perspective. I showed that a behavioural marker previously thought to index rule-based processes involving symbolic representations in humans was equally predictive of the behaviour of artificial agents not equipped with such features. In the last experiment, I investigated how humans learn simple problems and showed that they might have an innate predisposition to search for abstract rules, allowing them to generalise their knowledge to new contexts. Taken together, it is hoped that the work presented in this thesis furthers our understanding of the principles behind flexible, intelligent behaviour and sheds new light on their mechanistic implementation in neural circuits. |
| spellingShingle | Cognitive neuroscience Computational neuroscience Piwek, EP Neural mechanisms for flexible behaviour in humans and artificial neural networks |
| title | Neural mechanisms for flexible behaviour in humans and artificial neural networks |
| title_full | Neural mechanisms for flexible behaviour in humans and artificial neural networks |
| title_fullStr | Neural mechanisms for flexible behaviour in humans and artificial neural networks |
| title_full_unstemmed | Neural mechanisms for flexible behaviour in humans and artificial neural networks |
| title_short | Neural mechanisms for flexible behaviour in humans and artificial neural networks |
| title_sort | neural mechanisms for flexible behaviour in humans and artificial neural networks |
| topic | Cognitive neuroscience Computational neuroscience |
| work_keys_str_mv | AT piwekep neuralmechanismsforflexiblebehaviourinhumansandartificialneuralnetworks |