IIIT Hyderabad Publications
Large Scale modeling of the Basal Ganglia
Author: Bhargav Teja Nallapu
Report no: IIIT/TH/2016/68
Advisor:Bapi Raju Surampudi
Computational modeling helps us study the mechanisms underlying different cognitive and neural processes and relate them to the corresponding experimental data. Besides relating to the experimental data, a computational model should be able to propose a testable hypothesis. This thesis attempts to develop a computational model of a group of sub-cortical nuclei in the brain, called the Basal Ganglia (BG), to understand their role in reward based action selection and decision making. One of the exist-ing biologically plausible models of the BG is simulated and the model is able to learn a probabilistic two-armed bandit task using reinforcement learning. This model is able to choose the best option and reach optimal performances after only a few trials. Further, an extended decision making framework involving the BG is developed, accommodating a variety of stimulus representations and realistic scenarios. The framework is used to study the influ- ence of exogenous factors such as stimulus salience and delay in stimulus presentation. Analysis of the results suggests that a reward based action selection model, which performs optimally otherwise, performs sub-optimally when the visual characteristics of the stimuli such as salience are changed and when delays are introduced in the successive presentation of stimuli. The influence of salience changes and delays of presentation of stimuli on decision making form novel testable predictions from the com-putational experiments reported here. The developed framework is quite simplistic with as few as 72 neurons representing all the brain structures involved in decision making. It has been successfully expanded to a large scale both in terms of neuron populations and extending to a variety of tasks that the BG are known to be involved in. Scaling up involved exploring various large scale neural modeling frameworks like DANA 1 and NEF 2 (Nengo). The model is scaled using DANA to as many as 66,000 neurons. With the large scale model, the functional properties and neural activity patterns have been found to be consistent with the simplistic model. Scaling up using the large scale neural modeling platform, Nengo, also yielded the same dynamics of the simplistic model but now the model is expanded to a massive scale of 300K neurons. The power of converting existing models to efficiently weighed neural networks in Nengo is exploited in this work. Challenges involving scaling simple models, the dynamics while connecting largely and differently populated structures and pros and cons involved in using various modeling frameworks are discussed. Crucial aspects in a computational model, like parameter tuning and detailed neural implementations, while moving from a simplistic to large-scale model, are also studied. While the focus of the thesis has been to match the qualitative behavior at small and large scales, such large scale models have intrinsic advantages such as the ability to implement sophisticated stimulus representation, graceful degradation of performance in the face of injury / insult to the neural substrate and implementation of a biologically plausible architecture that matches details of the anatomy.
Full thesis: pdf
Centre for Cognitive Science
Copyright © 2009 - IIIT Hyderabad. All Rights Reserved.