Manjusha Nair

Efficient simulations of spiking neurons on parallel and distributed platforms: Towards large-scale modeling in computational neuroscience

Human brain communicates information by means of electro-chemical reactions and processes it in a parallel, distributed manner. Computational models of neurons at different levels of details are used in order to make predictions for physiological dysfunctions. Advances in the field of brain simulations and brain computer interfaces have increased the complexity of this modeling process. With a focus to build large-scale detailed networks, we used high performance computing techniques to model and simulate the granular layer of the cerebellum. Neuronal firing patterns of cerebellar granule neurons were modeled using two mathematical models Hodgkin-Huxley (HH) and Adaptive Exponential Leaky Integrate and Fire (AdEx). The performance efficiency of these modeled neurons was tested against a detailed multi-compartmental model of the granule cell. We compared different schemes suitable for large scale simulations of cerebellar networks. Large networks of neurons were constructed and simulated. Graphic Processing Units (GPU) was employed in the pleasantly parallel implementation while Message Passing Interface (MPI) was used in the distributed computing approach. This allowed to explore constraints of different parallel architectures and to efficiently load balance the tasks by maximally utilizing the available resources. For small scale networks, the observed absolute speedup was 6X in an MPI based approach with 32 processors while GPUs gave 10X performance gain compared to a single CPU implementation. In large networks, GPUs gave approximately 5X performance gain in processing time compared to the MPI implementation. The results enabled us to choose parallelization schemes suitable for large-scale simulations of cerebellar circuits. We are currently extending the network model based on large scale simulations evaluated in this paper and using a hybrid - heterogeneous MPI based multi-GPU architecture for incorporating millions of cerebellar neurons for assessing physiological disorders in such circuits.

GPGPU Implementation of a Spiking Neuronal Circuit Performing Sparse Recoding

Modeling and simulation techniques have been used extensively to study the complexities of brain circuits. Simulations of bio-realistic networks consisting of large number of neurons require massive computational power when they are designed to provide real-time responses in millisecond scale. A network model of cerebellar granular layer was developed and simulated here on Graphic Processing Units (GPU) which delivered a high compute capacity at low cost. We used a mathematical model namely, Adaptive Exponential leaky integrate-and-fire (AdEx) equations to model the different types of neurons in the cerebellum. The hypothesis relating spatiotemporal information processing in the input layer of the cerebellum and its relations to sparse activation of cell clusters was evaluated. The main goal of this paper was to understand the computational efficiency and scalability issues while implementing a large-scale microcircuit consisting of millions of neurons and synapses. The results suggest efficient scale-up based on pleasantly parallel modes of operations allows simulations of large-scale spiking network models for cerebellum-like network circuits.

Computational Neuroscience of Timing, Plasticity and Function in Cerebellum Microcircuits

Cerebellum has been known to show homogeneity in circuit organization and hence the “modules” or various circuits in the cerebellum are attributed to the diversity of functions such as timing, pattern recognition, movement planning and dysfunctions such as ataxia related to the cerebellum. Ataxia-like conditions, induced by intrinsic excitability changes, disable spiking or bursts and thereby limit the quanta of downstream information. Understanding timing, plasticity and functional roles of cerebellum involve large-scale and microcircuit reconstructions validating molecular mechanisms in population activity. Using mathematical modelling, we attempted to reconstruct information transmission at the granular layer of the cerebellum, a circuit whose role in dysfunctions remain yet to be fully explored. We have employed spiking models to reconstruct timing roles and detailed biophysical models for extracellular activity and local field population response. The roles of inhibition, induced plasticity and their implications in information transmission were evaluated. Modulatory roles of Golgi inhibition and pattern abstraction via optimal storage were estimated. An abstraction of the granular and Purkinje layer circuit for neurorobotic roles such as pattern recognition and spike encoding via two new methods was developed. Simulations suggest plasticity at cerebellar relays may be an important element of tremendous storage capacity reliable in the learning of coordination of actions, sensorimotor or cognitive, in which the cerebellum participates.

Modeling basal ganglia microcircuits using spiking neurons

Basal ganglia and cerebellum have been implicated in critical roles related to control of voluntary motor movements for action selection and cognition. Basal ganglia primarily receive inputs from cortical areas as well as thalamic regions, and their functional architecture is parallel in nature which link several brain regions like cortex and thalamus. Striatum, substantia nigra, pallidum form different neuronal populations in basal ganglia circuit which were functionally distinct supporting sensorimotor, cognitive and emotional-motivational brain functions. In this paper, we have modelled and simulated basal ganglia neurons as well as basal ganglia circuit using integrate and fire neurons. Firing behaviour of subthalamic nucleus and global pallidus externa show how they modulate spike transmission in the circuit and could be used to model circuit dysfunctions in Parkinsons disease.

Parameter optimization and nonlinear fitting for computational models in neuroscience on GPGPUs

One of the main challenges in computational modeling of neurons is to reproduce the realistic behaviour of the neurons of the brain under different behavioural conditions. Fitting electrophysiological data to computational models is required to validate model function and test predictions. Various tools and algorithms exist to fit the spike train recorded from neurons to computational models. All these require huge computational power and time to produce biologically feasible results. Large network models rely on the single neuron models to reproduce population activity. A stochastic optimization technique called Particle Swam Optimisation (PSO) was used here to fit spiking neuron model called Adaptive Exponential Leaky Integrate and Fire (AdEx) model to the firing patterns of different types of neurons in the granular layer of the cerebellum. Tuning a network of different types of spiking neurons is computationally intensive, and hence we used Graphic Processing Units (GPU) to run the parameter optimisation of AdEx using PSO. Using the basic principles of swam intelligence, we could optimize the n-dimensional space search of the parameters of the spiking neuron model. The results were significant and we observed a 15X performance in GPU when compared to CPU. We analysed the accuracy of the optimization process with the increase in width of the search space and tuned the PSO algorithm to suit the particular problem domain. This work has promising roles towards applied modeling and can be extended to many other disciplines of model-based predictions.

Temporal constrained objects for modelling neuronal dynamics

Background: Several new programming languages and technologies have emerged in the past few decades in order to ease the task of modelling complex systems. Modelling the dynamics of complex systems requires various levels of abstractions and reductive measures in representing the underlying behaviour. This also often requires making a trade-off between how realistic a model should be in order to address the scientific questions of interest and the computational tractability of the model. Methods: In this paper, we propose a novel programming paradigm, called temporal constrained objects, which facilitates a principled approach to modelling complex dynamical systems. Temporal constrained objects are an extension of constrained objects with a focus on the analysis and prediction of the dynamic behaviour of a system. The structural aspects of a neuronal system are represented using objects, as in object-oriented languages, while the dynamic behaviour of neurons and synapses are modelled using declarative temporal constraints. Computation in this paradigm is a process of constraint satisfaction within a time-based simulation. Results: We identified the feasibility and practicality in automatically mapping different kinds of neuron and synapse models to the constraints of temporal constrained objects. Simple neuronal networks were modelled by composing circuit components, implicitly satisfying the internal constraints of each component and interface constraints of the composition. Simulations show that temporal constrained objects provide significant conciseness in the formulation of these models. The underlying computational engine employed here automatically finds the solutions to the problems stated, reducing the code for modelling and simulation control. All examples reported in this paper have been programmed and successfully tested using the prototype language called TCOB. The code along with the programming environment are available at http://github.com/compneuro/ TCOB_Neuron. Discussion: Temporal constrained objects provide powerful capabilities for modelling the structural and dynamic aspects of neural systems. Capabilities of the constraint programming paradigm, such as declarative specification, the ability to express partial information and non-directionality, and capabilities of the object-oriented paradigm especially aggregation and inheritance, make this paradigm the right candidate for complex systems and computational modelling studies. With the advent of multi-core How to cite this article Nair et al. (2018), Temporal constrained objects for modelling neuronal dynamics. PeerJ Comput. Sci. 4:e159; DOI 10.7717/peerj-cs.159 Submitted 2 March 2018 Accepted 26 June 2018 Published 23 July 2018 Corresponding author Shyam Diwakar, shyam@amrita.edu Academic editor Nicolas Rougier Additional Information and Declarations can be found on page 27 DOI 10.7717/peerj-cs.159 Copyright 2018 Nair et al. Distributed under Creative Commons CC-BY 4.0 parallel computer architectures and techniques or parallel constraint-solving, the paradigm of temporal constrained objects lends itself to highly efficient execution which is necessary for modelling and simulation of large brain circuits. Subjects Computational Biology, Scientific Computing and Simulation, Programming Languages

Information theoretic visualization of spiking neural networks

Visualization is a flexible way to analyze simulated data and serves as a means for scientific discovery. Large scale neural simulations using high performance and distributed computing techniques produce huge amount of data for which visual analysis is generally difficult to perform. In this paper, a spiking neuron simulation environment was created to model and simulate networks of neurons of the cerebellum. Traditional visualization techniques were used to highlight relevant findings from small scale cerebellar networks. Time varying volume visualization using traditional techniques was found infeasible as network size increased. New data abstractions were required to depict the data that changes over time. With large scale cerebellar networks, Information theoretic methods were used to reduce dimensionality and to extract valuable information from data. We suggested that, information theory can be used as an efficient scientific data analysis and visualization tool to evaluate and validate computational models of cerebellar like structures.

Comparing parallel simulation of single and multi-compartmental spiking neuron models using gpgpu

Characterizing neural responses and behavior require large scale simulation of brain circuits. Spatio-temporal information processing in large scale neural simulations often require compromises between computing resources and realistic details to be represented. In this work, we compared the implementations of point neuron models and biophysically detailed neuron models on serial and parallel hardware. GPGPU like architectures provide improved run time performance for multi compartmental Hodgkin-Huxley (HH) type neurons in a computationally cost effective manner. Single compartmental Adaptive Exponential Integrate and Fire (AdEx) model implementations, both in CPU and GPU outperformed embarrassingly parallel implementation of multi compartmental HH neurons. Run time gain of CPU implementation of AdEx cluster was approximately 10 fold compared to the GPU implementation of 10-compartmental HH neurons. GPU run time gain for Adex against GPU run time gain for HH was around 35 fold. The results suggested that careful selection of the neural model, capable enough to represent the level of details expected, is a significant parameter for large scale neural simulations.

Computational reconstruction of fMRI-BOLD from neural activity

In this paper, we model function magnetic resonance imaging signals generated by neural activity (fMRI). fMRI measures changes in metabolic oxygen in blood in brain circuits based on changes in biophysical factors like concentration of total cerebral blood flow, oxy-hemoglobin and deoxy-hemoglobin content. A modified version of the Windkessel model by incorporating compliance has been used with a balloon model to generate cerebellar granular layer and visual cortex blood oxygen-level dependent (BOLD) responses. Spike raster patterns were adapted from a biophysical granular layer model as input. The model fits volume changes in blood flow to predict the BOLD responses in the cerebellum granular layer and in visual cortex. As a comparison, we tested the balloon model and the modified Windkessel model with the mathematically reconstructed BOLD response under the same input condition. Delayed compliance contributed to BOLD signal and reconstructed signals were compared to experimental measurements indicating the usability of the approach. The current study allows to correlate dynamic changes of flow and oxygenation during brain activation which connects single neuron and network activity to clinical measurements.

Information processing via post-synaptic EPSP-spike complex and model-based predictions of induced changes during plasticity in cerebellar granular neuron

Understanding functional role of spike bursts in the brain circuits is vital in analyzing coding of sensory information. Information coding in neurons or brain cells happen as spikes or action potentials and excitatory post-synaptic potentials (EPSPs). Information transmission at the Mossy fiber- Granule cell synaptic relay is crucial to understand mechanisms of signal coding in the cerebellum. We analyzed spiking in granule cells via a detailed computational model and computed the spiking-potentiation contributing to signal recoding in granular layer. Plasticity is simulated in the granule cell model by changing the intrinsic excitability and release probability of the cells. Excitatory post synaptic potentials and spikes on varying Golgi cell (GoC) inhibition and Mossy fiber(MF) excitation were analyzed simultaneously with the effect of induced plasticity changes based on the timing and amplitude of the postsynaptic signals. It is found that a set of EPSPs reaching maximum threshold amplitude are converted to less number of high amplitude EPSPs or spikes. Exploring the EPSP-spike complex in granular neurons reveal possible mechanisms and quantification of information encoding in individual neurons of the cerebellar granular layer. Therefore, our study is potentially an important estimation of cerebellar function.