Posters

P2. Lossless Compression of EEG Signals through modified Multivariate Autoregression algorithm implemented on FPGA

Md. Mushfiqur Rahman Chowdhury, Shubhajit Roy Chowdhury

Biomedical Systems Laboratory, School of Computing and Electrical Engineering, Indian Institute of Technology Mandi

Abstract: This paper reports of an efficient lossless compression of EEG signals based on modified Multivariate Autoregression algorithm implemented on Field Programmable Gate Array. The Multivariate Auto-regression (MVAR) algorithm has been refined by using least square matrix instead of using co-efficient matrix for prediction of electrode signals and it gave us higher compression ratio averaging more than 76.2% and, in some cases, more than 95%. The power consumption of the circuit is 0.511 micro-watts. The data is represented in lesser number of bits, the main data is used to predict data and determine the error signals which are sent to the receiver, in the receiving end the data is reconstructed by adding the predicted signal to the error signal. In this way, the original signal of the electrodes is compressed without any loss of information. Thus, we are able to reduce the transmission cost and increase transmission efficiency. In this research work, we have implemented FIR filter as EEG data compressor and data predictor circuit, the MVAR model has also been implemented in hardware by using Verilog HDL for the first time ever. The future task is to do live EEG data acquisition and compression by using this hardware implementation. The circuit which has been designed works for 22 channels. The same design can be replicated to create the circuit design for compressing 128 channels or 256 channels EEG data. The circuit has been implemented in Zynq Ultra scale ZCU 104 FPGA (Field Programmable Gate Array).

P4. A framework for Oscillatory Neural Network using Chaotic Oscillators

N.R.Rohan, Sayan Ghosh, Sayan Gupta, V Srinivasa Chakravarthy

CNS Lab, Department of Biotechnology, IIT Madras

Abstract: In this study we propose a new Oscillatory Neural Network framework using a single Chaotic Oscillator that is capable of modelling any N-channel EEG time series with high accuracy; particularly in the case of epileptic seizures. In a previous work this feat was achieved by a significantly large reservoir of Hopf Oscillators. Additionally, we explore four different systems that exhibit chaos and their efficacy in achieving high accuracy and efficiency while modelling an EEG time series.

P6. A comprehensive multi-scale computational model for Lower Limb Spasticity

Madhav Vinod Pithapuram

Indian Institute of Technology Hyderabad

Abstract: Conventionally scientists as well as clinicians have relied on animal models to study the movement and related disorders. In order to observe the internal state of motor machinery and underlying mechanism pertaining to movement or diseases, the research in humans is mostly limited due to dearth of non-invasive approaches and ethical consideration. The technology for experimentation has advanced and the results from animal experiments frequently raise many hypotheses, which are currently usually impossible to be tested on human. (Appenteng & Prochazka, 1984; Blaschak et al., 1988; Burke et al., 1972; Cleland & Rymer, 1990a; Duysens & van de Crommert, 1998; Fleshman et al., 1988; Hultborn & Pierrot‐Deseilligny, 1979; Ivashko et al., 2003; Jankowska, 2008a, 2008b; Jones & Bawa, 1999; Prochazka & Gorassini, 1998a, 1998b; Ranck & Bement~, 1965; Toossi et al., 2021). Computer simulations can help in the hypothesis testing by providing several procedures and measurements that is not available experimentally. Simulations can also be useful for theoretical investigations of neural systems properties, for demonstrating rules and paradigms in neuroscience and for the study of neuropathy.(Cisi & Kohn, 2008a; Falisse et al., 2020; M. Hines, 1989; Markram et al., 2015; Moore et al., 1978; Powers et al., 2012; Song & Geyer, 2018; Szlavik, 2008; Thelen et al., 2003). While biomechanics of movement are observable and recordable at high resolution, the internal states facilitating such movement are not measurable. For example, the neural controllers of the muscular motors, their configurations, states are not easily observable in humans, primarily due to lack of non-invasive nature of tools and limited methods currently available at hand. Our effort aims to cater this need by designing a pipeline/ framework to simulate these non-observable internal states of neuro-motor system at the level of spinal cord and musculoskeletal system which could be used in treatment design of movement disorders such as spasticity. The broad goal of this study is to understand the action of spinal cord neural circuits as a dynamic time series during spasticity. It relies on building detailed computational models of the spinal cord circuitry and their efferents/afferents to/ from various muscles and incorporating various internal and external forces acting on musculoskeletal system to demonstrate and verify the underlying neural mechanisms of spasticity in-silico.

P8. Modeling of whole brain Electroencephalogram (EEG) in a spatially organized oscillatory network

Sayan Ghosh; Dipayan Biswas; Sujith Vijayan; VS Chkravarthy

CNS Lab, Department of Biotechnology, IIT Madras

Abstract: In this study, we have introduced the spatial localization of oscillators in our existing complex oscillatory network. We have tried to model a large-scale TVB (The Virtual Brain) kind of network. Here two kinds of cortical sheets: rectangular and spherical, are described. In this study, we have introduced local or short-range connectivity among oscillators which basically describes the functional connectivity of the brain. Real 10-20 electrode geometry was also introduced while the spherical oscillatory model. In this study, lateral connection and natural frequency of oscillators were shared by two nearest channels. Ultimately, we have compared the model predicted signal with the original EEG signal in terms of time series as well as in the frequency domain.

P10. Predictive Coding in Auditory Cortical Neurons in Songbirds

Srihita Rudraraju; Michael Turvey; Brad Theilman; Timothy Gentner

UC San Diego

Abstract: Sensory cortex forms predictions about future events by integrating sensory information from the environment for efficient processing. This notion is hypothesized by predictive coding (PC), a theoretical framework in which the brain compares a generative model to incoming sensory signals. There is little understanding, however, of how it might be implemented at a mechanistic level in individual neurons within the auditory system. Here, we examined responses of single neurons in caudomedial nidopallium (NCM) and caudal mesopallium (CM), analogs of auditory cortex, in anesthetized European starlings listening to conspecific songs. We trained a feedforward temporal prediction model (TPM) to define a “latent” predictive feature space and its corresponding feature space representing prediction error. We show that NCM responses are best modeled by the predictive features of spectrotemporal song, while CM responses employ both predictive and error features. This provides strong support for the notion of a feature-based predictive auditory code implemented in single neurons in songbirds..

P12. Biologically plausible model to recogne structure from motion and biological motion from a stimuli where an actor is embedded in a rotating shape

Anila Gundavarapu, V. Srinivasa Chakravarthy

CNS Lab, Department of Biotechnology, IIT Madras

Abstract: Our visual system can understand actions performed by others or can recognize objects around us only from sparse motion cues such as (i) point light (PL) displays (ii) rotating random-dot surfaces, phenomena called biological motion processing (BM), and structure from motion (SFM) respectively. To understand the mechanisms involved in the above phenomenon we developed a hierarchical structure from motion network (SFMNW) and biological motion network (BMNW). SFMNW is trained to recognize the shape of a rotating 3D surface and BMNW is trained to recognize action performed by a PL display. The trained two models are presented with mixed stimuli (a point light action embedded in a rotating shape) and achieved 100% accuracy in shape recognition and 98% accuracy in action recognition. Through our simulation, we presume that the robust motion circuitry that supports SFM processing is also involved in the BM processing instead reliable SFM perception might involve a contribution from ventral areas too. The proposed models provide new insight into the research on the development of two phenomena SFM and BM processing.

P14. A Comparative Analysis for a newly designed trainable oscillatory Neural Network, and Basic neural models

Anirban Bandyopadhyay1; Sayan Ghosh1; Dipayan Biswas1; Raju Bapi Surampudi2 ; V. Srinivasa Chakravarthy1

1.Computational Neuroscience Lab, Dept.- Biotechnology,Indian Institute of Technology Madras

2.Brain,Cognition, and Computation Lab, International Institute of Information Technology Hyderabad

Abstract: Abstract level whole brain dynamical model has become instrumental to capture the whole brain neuroimaging signals, like BOLD signals. Newly developed hopf-oscillator based biophysical model comprises of two phases- in the first phase the oscillators positioned in the corresponding ROIs (Region of Interest), have a lateral connection with each other, and with a Fourier like decomposition, the frequencies and each angle of the power coupling has been learnt; in the second phase the oscillators with previously learnt parameters are trained in a hidden layer of thirty neurons-based feedforward network for each ROI. The connections between oscillators are determined by the structural connectivity. After 20,000 epochs, the predictive power (correlation coefficient) between the simulated and empirical functional connectivity matrix is calculated (Mean - 0.96, Std. Deviation- 0.0112 for ten healthy participants), which is far greater than the predictive power derived from the basic neuron models. [Reference for basic models - https://pubmed.ncbi.nlm.nih.gov/25682944] .

P16. Multi-processing nature of Human vision and perception models

Abhishek Tiwari

IIT Madras

Abstract: With recent advances in Deep Learning, Convolution Neural Networks (CNNs) have achieved even more accuracy than humans (more than 95%) on image classification problems. However, human vision is more than just a classification model, and high accuracy does not suggest that human brains have a similar neural network as CNNs. Research shows that CNNs focus on local features or shape instead of global features or shape as a whole. They also seem to focus on textures, and CNNs can learn to classify unstructured data, which a human brain will find almost impossible. These results suggest that the human brain multi-processes image content's global shape, spatial arrangement, and colour.

Here proposition is that humans classify image contents based on their global shape by multi-processing the detected edges, texture, and colour separately to identify image class rather than taking pixel-by-pixel data to focus on local features.

P18. A Comprehensive Deep Learning Network to Model Object Encoding in Hippocampal Formation

Azra Aziz, Bharat K. Patil, V Srinivasa Chakravarthy

Indian Institute of Technology Madras

Abstract: Many spatial cells are discovered to make mental spatial maps that help us navigate collectively. Many experimental studies conclude the importance of vision and proprioception to locate oneself in the environment accurately. For navigation purposes, the leading information we use from vision is landmarks and objects. Although a wide variety of models are proposed for grid cells and place cells as proprioception is enough to get these kinds of responses, there are no models for the neurons that respond to landmarks and objects. We propose a comprehensive deep network that combines realistic vision from Virtual reality (VR) with proprioception to model a wide variety of neurons that respond to objects and landmarks in different ways, along with grid cells and place cells. Here we mainly focus on object-sensitive cells in the lower entorhinal cortex (LEC), object vector cells in the medial entorhinal cortex (MEC), and landmark vector cells in the CA1 region. The model comprises two parallel pathways: 1) vision pathway through convolution layers and 2) path integration pathway. The two pieces of information communicate at a higher level through graph neural networks (GNN), and the network is trained on a reward scheme. We observe the desired results in the hidden layers of the model.

P20. Learning Spatial Prepositions “far” and “near” using Convolution Neural Networks and Synthetic 3D Images

Krishna Raj S R 1,Dr. Anindita Sahoo 1 and Dr. V Srinivasa Chakravarthy2

1 :Department of Humanities and Social Sciences, Indian Institute of Technology Madras, Chennai, 600036, TamilNadu, India.

2: Department of Biotechnology, Indian Institute of Technology Madras, Chennai, 600036, TamilNadu, India

Abstract: Human language and semantics are strongly colored by our sensory-motor experiences. The constraints offered by the spatiotemporal world in which we live and manipulate shape the development of language syntax. Sensory experiences gathered over a background of a Spatio-temporal world are used as the raw material to create more abstract concepts. For example, when we speak of gaining a “deep understanding” of a subject, we are borrowing a spatial concept and superimposing it on an abstract concept. Studying spatial prepositions and vision helps gain insights into the interplay of space and language.

We hereby present a computational study that investigates the influence of the various properties of the sensory system on the acquisition of a select set of spatial prepositions. A synthetic image dataset is created for the study. A Convolution Neural Network (CNN) based network architecture is trained to distinguish between two prepositions ‘far’ and ‘near’ which represent the relationship between objects in the image. The results indicate that the presence of extra objects (perspective aids) and the texture of the extra objects affect the network's performance. Only when the objects are confined to a plane the network was able to distinguish between the prepositional relationship 'far' and 'near'.

P22. Spectral graph modeling of Alzheimer’s disease neurophysiology

Parul Verma (a) , Kamalini G Ranasinghe (b) , Chang Cai a , Xihe Xie (a) , Hannah Lerner b , Danielle Mizuiri (a) , Bruce L

Miller (b) , Katherine P Rankin (b) , Keith A Vossel (b,c) , Srikantan S Nagarajan (a*) , Ashish Raj (a*)

a ) Department of Radiology and Biomedical Imaging, University of California San Francisco, San Francisco,

CA

b) Memory and Aging Center, Department of Neurology, University of California San Francisco, San Francisco,

CA

c) Mary S. Easton Center for Alzheimer’s Disease Research, Department of Neurology, David Geffen School of

Medicine, University of California Los Angeles, Los Angeles, CA 90095

* Co-corresponding author

Abstract: Alzheimer's disease (AD) is the most common form of dementia, progressively impairing memory, cognition, as well as behavior. While various neuroimaging studies of the human brain have revealed functional abnormalities in patients with AD [1], how neuronal and synaptic functions are impaired remain unclear. Electrophysiological recordings that capture the local field potentials from pyramidal neuronal firing are the most direct techniques to record neuronal activity from human subjects non-invasively. Combined with mathematical models, empirical spectral data from electrophysiology will help uncover the abnormal biophysical mechanisms of neuronal activity which is otherwise intractable. In this work, we employed a spectral graph-theory based model (SGM) to identify abnormal biophysical markers of neuronal activity in AD [2].

SGM is a hierarchical and analytic model that describes the dynamics of excitatory and inhibitory neuronal activity. It models the coupled excitatory and inhibitory activity of local neuronal subpopulations, and the long-range excitatory macroscopic dynamics, for every brain region. It is parameterized by a small set of global parameters – we inferred these parameters for a well characterized clinical population of AD patients and a cohort of age-matched controls [3]. We estimated model parameters that best captured the regional MEG frequency spectra as well as the spatial distribution of the empirical alpha frequency oscillation. For both patients and controls, the modeled spectra closely match the empirical MEG spectra (Fig 1A). Patients with AD have significantly elevated long-range excitatory neuronal time constant () compared to controls (p = 0.0006; Fig 1B). Moreover, higher is also associated with cognitive deficits in AD.

Specifically, higher is positively correlated with Clinical Dementia Rating – Sum of Boxes (p =0.0117; Fig 1C), and is negatively correlated with Mini Mental State Exam score (p = 0.0016; Fig 1D). These results indicate that abnormal spectral signatures in combination with SGM can reliably depict altered excitatory neuronal activity in AD patients. Importantly these abnormalities showed significant associations with cognitive deficits. These findings are intriguing given the context that neurofibrillary tangle pathology that is closely allied to cognitive deficits in AD have been shown to preferentially affect excitatory neurons in human neuropathological studies [4,5]. Our findings provide critical insights about potential mechanistic links between abnormal neural oscillations and cellular correlates of impaired neuronal activity in AD.

P24. Impaired Arbitration Between Decision-Making Strategies In Alcohol Users: A Computational Modeling Study

Srinivasan Ramakrishnan 1,2 , Jeffrey Chen 2,3 , Gopi K. Neppala 2 , Riaz B. Shaik 2 ,

Wouter Kool 4 , Iliyan Ivanov 2 , Muhammad A. Parvaz 2,5

Abstract: Studies of reinforcement learning propose that decision-making is guided by a tradeoff between computationally cheap model-free control (habit) and costly model-based (goal- directed) control. Indeed, humans depend on model-based control when higher rewards are at stake. Addicted individuals purportedly rely more on model-free control; however, it remains to be studied how they arbitrate between model-free and model-based control under higher stakes of reward.