Learning and cognition in brain and machine: Prediction of dementia from longitudinal data and modelling memory networks
MetadataVis full innførsel
OriginalversjonMofrad, S. A. (2021). Learning and cognition in brain and machine: Prediction of dementia from longitudinal data and modelling memory networks [Doctoral dissertation]. Western Norway University of Applied Sciences.
Starting in the mid-20th century and throughout their developments, modern neuroscience and artificial intelligence (AI) have provided each other with inspiration, insights, and tools. The degree to which they are intertwined has been in constant flux over the years, but always present. With the enormous resurgence of interest in machine learning over the past decade, led by the much-celebrated successes of artificial neural networks and deep learning, the bond between the two fields seems to be growing stronger. Artificial intelligence and machine learning have always kept an eye on biological intelligence and learning, as these provide our only examples of general intelligence and strong learning capabilities, inspiring the development of their much less capable–albeit improving–counterparts, which are based on computational models. The growing attention to both neuroscience and AI is also leading to growth where they intersect, i.e. in neuroscience-inspired AI and AI-inspired neuroscience, and in the usage of computational AI models within neuroscience and the cognitive sciences. In this context, the present thesis aims to make a modest contribution through our application of machine learning techniques to the study of dementia using data from longitudinal MRI and psychometric testing, and through our proposed models for simulating aspects of the formation of memory networks during learning and memory retrieval. The former is our main contribution and is addressed in studies A, B, and C, while the latter is reflected in studies D and E. Through longitudinal studies, i.e. studies based on the collection of repeated measurements from the same subjects, or experimental units, over time one can observe how measurements develop and discover new relationships between variables. Longitudinal data analysis is a large field of research comprised of a multitude of methods and is widely applicable to e.g. behavioural analysis and medicine. One inherently longitudinal phenomenon of particular interest for the present work is the biological, neurological, and cognitive alteration linked to aging. There is an immense need to develop methods that can indicate the risk of developing aging-related diseases such as dementia, as well as for increasing the understanding that is derived from new computational models for cognitive skills such as memory and learning. The first part of this thesis (studies A, B, and C) develops and evaluates methods for using machine learning models with longitudinal data that have a time-dependent structure. We propose two novel and flexible frameworks to describe the trajectories of change extracted from the longitudinal data. The two frameworks are, respectively, based on (i) a combination of mixed-effects models in order to extract features from the longitudinal trajectories that can be used to train any type of machine learning classifier and (ii) mapping the multi-dimensional data onto two-dimensional images, enabling classifications based on convolutional neural networks. The second part of this thesis (studies D and E) aims to construct simple and flexible models that can be used to simulate learning and memory retrieval processes in the human brain. These proposed memory networks are: (i) defining a new associative memory for storing sequences and investigate how to make efficient retrievals, and (ii) a combination of a reinforcement learning model to form memory connections in the training phase and an iterative diffusion process to update the memory network to be used in the test phase. We found that the frameworks proposed in the first part of the thesis, although being relatively simple approaches to the complexities of longitudinal data analysis, are comparable to other approaches in the literature as regards accurately predicting dementia. The proposed model for learning and retrieval based on associative memory in Paper D has several features that make it resemble its biological brain counterpart more than comparable models in the literature do, while significantly reducing errors in sequence-retrieval. The model for episodic memory developed in Paper E is quite flexible and can provide simulations of actual experiments on typical and atypical human behaviours.
Papers D and E are © 2021 MIT Press. Reprinted by permission.
Består avMofrad, S. A., Lundervold, A., & Lundervold, A. S. (2021). A predictive framework based on brain volume trajectories enabling early detection of Alzheimer's disease. Computerized Medical Imaging and Graphics, 90. https://doi.org/10.1016/j.compmedimag.2021.101910
Mofrad, S. A., Lundervold, A. J., Vik, A., & Lundervold, A. S. (2021). Cognitive and MRI trajectories for prediction of Alzheimer’s disease. Scientific Reports, 11(1). https://doi.org/10.1038/s41598-020-78095-7
Mofrad, S. A., Bartsch, H., Lundervold, A. S., & Alzheimer’s Disease Neuroimaging Initiative. (2021). From longitudinal measurements to image classification: Application to longitudinal MRI in Alzheimer’s disease [Manuscript submitted for publication]. Western Norway University of Applied Sciences.
Mofrad, A. A., Mofrad, S. A., Yazidi, A., & Parker, M. G. (2021). On neural associative memory structures: Storage and retrieval of sequences in a chain of tournaments. Neural Computation, 33(9), 2550-2577. https://doi.org/10.1162/neco_a_01417
Mofrad, A. A., Yazidi, A., Mofrad, S. A., Hammer, H. L., & Arntzen, E. (2021). Enhanced equivalence projective simulation: A framework for modeling formation of stimulus equivalence classes. Neural Computation, 33(2), 483-527. https://doi.org/10.1162/neco_a_01346