Unsupervised Deep learning for Human Cognition
Lead Research Organisation:
University College London
Department Name: Experimental Psychology
Abstract
Background
A fundamental challenge in machine learning is building useful representations. The expressiveness of deep convolutional neural networks is well suited to building representations for complex problems such as object recognition. Recently, more attention has been drawn to the joint research of human cognition with deep learning approaches. In particular, deep neural networks are being applied to neural imaging data from, for example, fMRI, in order to study complex embedding spaces in the brain with a goal to uncover more hidden secrets of the brain such as its dimensional functionalities.
Research Questions
1. Do bilingual brains have functional cognitive difference from monolinguals'?
2. Can priming effect be detected through brain imaging techniques during a decision-making process?
3. Can deep learning approaches such as auto-encoders be used to uncover more meaningful neural embedding spaces based on brain imaging data?
Research Approaches
General research methodologies would involve a combination of technical computing, programming and simulation with designed psychological experiments and collecting big-data datasets through other channels.
1. Difficulties might arise from the fact that brain signals can be very noisy and complex. Thus by nature it is a very challenging task along with the fact that deep learning results are known for hard to interpret which makes it even harder to evaluate the compatibility of DL models with cognitive models and integrate them.
2. In terms of technical approaches, a number of architectures will be considered, such as generic auto-encoders, variational auto-encoders, SWWAE auto-encoders, "Siamese" networks, and adversarial variants. A number of innovative extensions to these basic approaches will be considered with the aim of creating embedding spaces that are more aligned with human intuitions.
3. Data will be sourced from numerous channels, which include for example real-world dataset on related problems, neural data collected through lab experiments etc.
4. There could be some ethical considerations when experiments involve human participants.
Values
A general aim of this proposed work is to ground these advances in AI to the social sciences by building models that mimic human understanding using data sources and methods from the social sciences. The proposed research has the potential to create value in the joint area of cognitive science and artificial intelligence, particular in uncovering more sophisticated understanding of the human brain in order to facilitate human-machine interactions and potential applications in creative industries.
Statement of Relevant Qualifications
I have a master's degree in MSc Information Science with experiences in machine/deep learning techniques. I have had experiences with the techniques that are proposed and currently be studied in relevant areas such as auto-encoders and other representation learning algorithms. I will need to take the AQM training.
Potential Publication
Top publication outlets will be targeted across fields, including proceedings for AI conferences (e.g., NIPS, ICLR, ICML), Psychology and Cognitive Science journals (e.g., Psychological Review, Cognitive Science, Cognition), and science general journals (e.g., PNAS, Journal of Royal Society Interface, Nature).
A fundamental challenge in machine learning is building useful representations. The expressiveness of deep convolutional neural networks is well suited to building representations for complex problems such as object recognition. Recently, more attention has been drawn to the joint research of human cognition with deep learning approaches. In particular, deep neural networks are being applied to neural imaging data from, for example, fMRI, in order to study complex embedding spaces in the brain with a goal to uncover more hidden secrets of the brain such as its dimensional functionalities.
Research Questions
1. Do bilingual brains have functional cognitive difference from monolinguals'?
2. Can priming effect be detected through brain imaging techniques during a decision-making process?
3. Can deep learning approaches such as auto-encoders be used to uncover more meaningful neural embedding spaces based on brain imaging data?
Research Approaches
General research methodologies would involve a combination of technical computing, programming and simulation with designed psychological experiments and collecting big-data datasets through other channels.
1. Difficulties might arise from the fact that brain signals can be very noisy and complex. Thus by nature it is a very challenging task along with the fact that deep learning results are known for hard to interpret which makes it even harder to evaluate the compatibility of DL models with cognitive models and integrate them.
2. In terms of technical approaches, a number of architectures will be considered, such as generic auto-encoders, variational auto-encoders, SWWAE auto-encoders, "Siamese" networks, and adversarial variants. A number of innovative extensions to these basic approaches will be considered with the aim of creating embedding spaces that are more aligned with human intuitions.
3. Data will be sourced from numerous channels, which include for example real-world dataset on related problems, neural data collected through lab experiments etc.
4. There could be some ethical considerations when experiments involve human participants.
Values
A general aim of this proposed work is to ground these advances in AI to the social sciences by building models that mimic human understanding using data sources and methods from the social sciences. The proposed research has the potential to create value in the joint area of cognitive science and artificial intelligence, particular in uncovering more sophisticated understanding of the human brain in order to facilitate human-machine interactions and potential applications in creative industries.
Statement of Relevant Qualifications
I have a master's degree in MSc Information Science with experiences in machine/deep learning techniques. I have had experiences with the techniques that are proposed and currently be studied in relevant areas such as auto-encoders and other representation learning algorithms. I will need to take the AQM training.
Potential Publication
Top publication outlets will be targeted across fields, including proceedings for AI conferences (e.g., NIPS, ICLR, ICML), Psychology and Cognitive Science journals (e.g., Psychological Review, Cognitive Science, Cognition), and science general journals (e.g., PNAS, Journal of Royal Society Interface, Nature).
Organisations
People |
ORCID iD |
| Bradley Love (Primary Supervisor) | |
| Xiaoliang Luo (Student) |