Abstracts
P1: Temporal Cascade of Frontal, Motor and Muscle processes during human action stopping
Sumitash Jana; Ricci Hannah; Vignesh Muralidharan & Adam R Aron
Aron Lab, Department of Psychology, University of California, San Diego, USA.
Abstract:
Rapid action stopping, a canonical executive function, is thought to be mediated by top-down fronto-basal-ganglia control over the motor system. Here, in humans, we map out the temporal progression of events in this putative fronto-BG-motor circuitry with unprecedented detail. We show that, following the signal to stop, there is an increase of right frontal beta at ~120 ms, a proxy of right inferior frontal gyrus; then, at 140 ms, there is a broad skeletomotor suppression, likely reflecting the impact of the subthalamic nucleus on BG output; then at ~160 ms, suppression is detected in the muscle, and, finally after a peripheral delay, the behavioral time of stopping (SSRT) is ~220 ms. These neural and muscle events, parse out the sub-processes in the fronto-BG-motor system at a temporal resolution, hitherto unknown. Furthermore, causally intervening the system around the time of the right frontal beta activity seems to affect stopping behavior.
P2: Generalized deep network model that recognize various types of optic flow
Abstract:
From the extensive research of the visual motion processing in the primates it is known the complex motion processing like optic flow is believed to be processed by the two areas MT and MST where neurons in MT contain many direction selective cells that in principle encode the flow field arriving on the retina. The neurons in MST have large receptive fields and respond to the whole optic flow patterns such as clock wise/ anti-clock wise rotation, zoom in or zoom out motion. Here we propose a feed forward deep network that receives input from neural field mosaic network, that can recognize the type of the optic flow hidden in the given image sequences. All the image sequences are made up of 2x2 tiny white squares placed on black background, and are allowed to move in translational, radial and rotational trajectories. Recognition accuracy obtained on test set is 90%.
P3: Spherical Harmonics for a basis for representing 3D space in a network of hippocampal spatial cells
1 Department of Chemical Engineering, Indian Institute of Technology Madras, Chennai, India;
2 Department of Biotechnology, Indian Institute of Technology Madras, Chennai, India;
3 Department of Psychology, University of California, Riverside, USA;
4 Department of Physics, Indian Institute of Technology Madras, Chennai, India.
Abstract:
We present a computational model of hippocampal spatial cells in 3D navigation. The network model has three layers. The head direction layer is a two-dimensional array of neurons representing preferred pitch and azimuth. The head direction responses are fed to a Path Integration (PI) layer which codes for displacement in the preferred direction from the initial point and also integrates multiple β which is a modulation factor. The final layer performs Principal Component Analysis (PCA) and extracts the Principal Components that provide a basis to represent the encoded spatial information. In 2D navigation, the principal components turned out to be sinusoids. As an analogous phenomenon, the PCs obtained in the 3D navigation model are Spherical Harmonics, which constitute the Fourier basis for a 3D space. This study reveals a plausible coordinate system that the brain uses to represent 3D space. The model further reveals place cell behavior in 3D.
P4: A Deep Learning Approach For 2D Spatial Cognition
1 Department of Biotechnology, Indian Institute of Technology Madras, Chennai, India;
2 Chemical Engineering department, Indian Institute of Technology Madras, Chennai, India;
3 Department of Electronics Engineering, Indian Institute of Technology (BHU) Varanasi, Varanasi, India;
4 Department of Physics, Indian Institute of Technology Madras, Chennai, India.
Abstract:
A computational model for 2D navigation is proposed for understanding spatial cognition. The first layer in model consists of head direction (HD) cells which codes for preferred direction of the agent. The second layer is path integration (PI) layer which calculates the displacement in the preferred direction from the initial point and also integrates multiple β which is a scaling factor for grid cell. This data is trained using stacked auto encoder and results are analyzed from the middle layer of every stack. The middle layer of first stack give grid cells. Similarly, the next stack middle layer gives place cells. Since different sensory information such as vision and audition are crucial for an agent to navigate more efficiently, therefore auto encoder model is been proposed as it is capable of handling non-linearity in the data for future work.
P5: A Systems Neuropharmacological Model of Cortico-basal ganglia circuitry of Arm Reaching for Normal, Parkinsonian and Levodopa Medication
Abstract:
To investigate the underlying mechanisms involved in Parkinson’s disease (PD) and its various symptoms, it is important to have an integrated computational model that links SNc cell loss to the behavioral symptoms. We developed a systems neuropharmacological model of cortico-basal ganglia motor circuitry for reaching task that simulates the behaviour associated with normal, PD, and medicated PD conditions. The initial results successfully demonstrated the PD on-off condition and also some of the cardinal symptoms of PD. This model can be expanded to a potential testbench for PD patients. The proposed model has a potential clinical application where drug dosage can be optimized as per patient characteristics. In future, we aim to simulate and characterize the effects of different levodopa medication anomalies, such as wearing off mechanism and levodopa induced dyskinesia.
P6: Detection system for Mudra Bharati - Finger-spelling for Indian scripts
Amal Jude Ashwin F1; Sunil Kumar Kopparapu2 & V Srinivasa Chakravarthy1
2 TCS Research and Innovation, Mumbai, India.
Abstract:
Sign Language is a potential tool for communication in the deaf and dumb society. As proper nouns cannot be gestured accurately at one shot, finger-spelling is adopted to spell out names and places. Due to rich vocabulary and diversity in Indian scripts, it is cumbersome to use the American Sign Language convention for finger-spelling in Indian languages. Therefore, we propose a novel convention for the finger-spelling system in Indian scripts, Mudra Bharati, whose dictionary is constructed based on the phonics of aksharas - 16 vowels and 40 consonants. Unlike ASL that utilizes just one hand, Mudra Bharati uses two hands- one for consonants and the other for vowels and samyukta aksharas are gestured by combining the vowel and the consonant. A prototype of the detection system for Mudra Bharati that returns the character in Devanagari and Tamil scripts is developed using Self-Organizing Maps and Convolutional Neural Networks.
P7: Cortical paired associative stimulation influences response inhibition by modulating cortico-subcortical networks
1 University of Cambridge, UK;
2 Kyoto University, Japan.
Abstract:
The ability to inhibit an ongoing action is critical to decision making. Response inhibition is applicable not only in daily functioning but its impairments can be observed across various psychiatric disorders. Cortical paired associative stimulation (cPAS) is a transcranial magnetic stimulation protocol, which involves repetitive low-frequency paired stimulation of two cortical regions which can either increase (long term potentiation-like) or decrease (depression-like) the synaptic strength between them, depending on the stimulation interval. PAS is commonly applied to the motor domain and not been tested extensively in the cognitive domain. A novel protocol to stimulate the right inferior frontal cortex (IFC) and pre-supplementary motor area (pre-SMA) at different intervals (4 milliseconds) was designed to show improvement in the stop signal task’s performance. A total of 34 healthy volunteers (aged 20-59 years) who were free from for any neurological or psychiatric conditions participated in a single session cPAS. The subjects completed the stop signal at baseline and after stimulation. We were able to successfully reproduce our earlier hypothesis and specifically show an improvement in the performance to be age dependent. There was a consistent improvement in the performance of the older subjects (p <0.05). These preliminary results validate our hypothesis that c-PAS can alter response inhibition circuitry through an age dependent synaptic mechanism, by modifying the connectivity strengths between cortico-cortico and cortico-sub cortical networks.
P8: Evaluating Exploration vs Exploitation for Stroke Rehabilitation: A Computational Study
Abstract:
Stroke is one of the leading causes of acquired disability in adults. Physiotherapy has shown great success in facilitating recovery. However, most rehabilitation techniques used are task-based and repetitive in nature and only cover a subset of movements preformed by a healthy human being in daily life. This leads to poor retention and generalization. Thus, in order to avoid this, it is essential to incorporate a wide variety of movements. To test this hypothesis, we propose a CNN model performing a simple visuomotor task. The input is composed of 2 images of a target (ball) as seen by the left and right eye and the output consists of muscle activation required to reach the ball. Stroke is induced in the model and rehabilitation is performed in two modes – exploration and exploitation. We observed that performance is much better with a more global effect after exploration compared to exploitation.
P9: Is Vascular Architecture Tailored to Bring Efficient Neural Performance? A Study of Artificial Neurovascular Network
Abstract:
Neural networks perform classification and functional-approximation tasks with reasonably high accuracy when trained optimally. However, unlike the neurons in artificial neural networks, biological neurons require energy to function. This energy is supplied by an extensive network of blood vessels. In this study we explore the effect of a trained vascular network on the neural network performance. The ANVN comprises a single layered MLP connected bidirectionally to the vascular tree structure. The root node of the vascular tree structure is an energy source and the terminal leaf nodes supply energy to the hidden neurons of the MLP. The branch weights depict the energy split from the parent node to the child node. The leaf node energies determine the bias of the hidden neuron. When analysing the test performance of ANVN for trained and untrained vascular networks, we found that higher performance is achieved for lower root-energies when the vascular network is trained.
P10: Attractor Dynamics in an Artificial Neurovascular Network
Abstract:
With increasing hidden layer size, a neural network shows increasing training performance while test performance reaches a maximum at an optimal size of the hidden layer. In this study the “energy efficiency” of an Artificial Neurovascular Network also reaches an optimum at a finite size of the hidden layer. This is because the increase in training accuracy is accompanied with an increase in total energy used. Can we find the optimal hidden layer size by observing the relation between energy consumption and accuracy for different hidden layer sizes? In this study, an MLP is connected bidirectionally to a simple vascular tree structure. An energy source supplies the energy demanded by the vascular tree. The biases of the hidden layer neurons are decided by the energy available at the leaf nodes near each neuron. The vascular weights are updated depending upon the requirement of each neuron. For a smaller value of hidden layer size, the network approaches a stable fixed point in the per capita energy consumption vs accuracy plot, whereas once the hidden layer size crosses a threshold, the fixed point appears to vanish giving place to a line of attractors.
P11: Convolutional Elman Jordan Neural Network in Action Recognition
Abstract:
Video processing has always been an important task in the field of computer vision. A lot of problems currently posed in this field requires recognition of actions in real time. Few examples of such problems are the detection of violent threats from in public places; classification of rogue vehicles on the road, etc. Solving one such problem can have a profound impact on society and make it a much safer place. Currently, these problems are being dealt with in a fashion that requires information present in the future. For example, one such architecture that does this, feeds a stack of frames from videos into multiple Convolutional Neural Networks present in an ensemble. Those CNNs extract spatio-temporal features from the data, which is then fed back into the system to make predictions. This study proposes a biologically inspired model, to classify actions in videos without making use of additional information that might not be present in real time. Image frames are extracted and fed into the network one by one, and the network makes predictions once enough images have been consumed. This studies that follow take this network and test its capabilities at classifying action based on the motion present.
P12: Addressing the working memory for searching task
Abstract:
Attention and memory have many possible forms of interaction. If memory has a limited capacity, it makes sense for the brain to be selective about what is allowed to enter it. In this way, the ability of attention to dynamically select a subset of total information is well-matched to the needs of the memory system. Inspiring from this, we have proposed a model which takes inputs of some patches of the big image to predict the class. Here patch refers to an attention window. Architecture of the model is designed with three pipelines in a way of predicting three outputs, one is location of the attention window of next time, one is the class of the big image, and the other is reconstruction of the big image. Network keeps storing the integrated features and locations of attention windows through time to make the classification prediction. Each pipeline consists of some convolutional layers or flip flop layers, maxpool layers and fully connected layers or flip flop layers. Flip flop layers have been used for storing the integrated information in memory.
P13: Analyzing the effect of trajectory and rotation parameter of camera on 3D object reconstruction
Abstract:
In this work, a 3d recurrent neural network is used to reconstruct 3d objects from 2d images. Parameters like the Z rotation of the camera and the trajectory of the camera are experimented with to find out the effect they have on the reconstruction of the 3d object. A data set is created which consists of the mass distribution of objects in 3d space and 2d images of the objects taken from various views. The network learns to map the input images of the object to the object’s 3d structure. The experiment shows that the Z rotation of the camera and the trajectory to be followed need not be fixed to reconstruct the 3d object.
P14: Bharati Script: A Simplified Common Script for Various Indian Languages
Abstract:
India being a land of multiple languages and scripts, the need for a simplified common script to reduce the communication barrier among people has been put forward by some intellectuals, especially during the post independence era. Bharati script is a step in this direction whose letters are carefully chosen so as to be easy not only for human beings but also for machines to read(/ interpret) and write. This is facilitated by a special 3 tier modular structure, that uses a minimal number of symbols(shapes) to express all letters of various Indian languages. Further as this script takes advantage of the phonetic commonalities of various letters(vowels/ consonants), it is easy to remember and that makes people feel like learning it. Once implemented,, this is sure to increase the literacy levels in the society and will help lakhs of non-resident Indian children too to read books in their mother tongue.
P15: Predictive Coding in the Auditory Cortex
Department of Psychology, University of California San Diego, USA.
Abstract:
Predictive Coding is forged by the brain using multiple layers of predictive assumptive models. Essentially, the brain receives input data from sensory stimulation, makes non-linear assumptions based on its previous knowledge of the world, and is quickly adjusted using corrected feedback by income data that follows. Prediction errors are used to select better top-down guesses, continually refining the predictive model, thus learning to minimize error for maximum optimization. In this study, we explore probabilistic generative models to construct plausible representation of sensory input called “prediction” and introduce a technique to test the predictive capacity of our generative model.
P16: Memory consolidation model of Prefrontal cortex for Lifelong learning
2 Department of Computer Science and Engineering & Robert Bosch Centre for Data Science and AI, Indian Institute of Technology Madras, Chennai, India.
Abstract:
Memory consolidation is the process of transferring the short-term memory from the network formed between the hippocampus and its associated cortical regions to the long-term memory network formed between the prefrontal cortex and its associated cortical structures through continuous replaying. This long-term memory stores the replayed memory along with already learned memory. Biologically inspired deep neural network models of this long-term storage suffers from catastrophic forgetting, in which the network forgets the old information upon learning the new information. In this work, we propose a model to solve the problem of catastrophic forgetting by controlling the direction in which parameters are updated. This is realized by adding an extra regularization-based loss function along with conventional loss function. The performance of the proposed model on retaining the old information is compared with Elastic Weight Consolidation on classification tasks and Autoencoder task.
P17: State anxiety differentiates value-driven and random exploration in an aversive multi-arm bandit experiment
National Institute of Mental Health/ National Institutes of Health, USA.
Abstract:
Background: Human decision making often involves exploration and exploitation, which may be impacted by anxiety. The goal of this study was to investigate the interaction between different exploration strategies (value-driven or random) and anxiety, using a computational model.
Methods: 32 healthy volunteers (18 F, mean age 25) performed 2 versions of a 4-armed aversive (probabilistic shock) bandit task, in which participants were asked to choose the arm perceived to have the lowest probability of shock. In one version, the best strategy relies on purely value-driven exploration (choosing to explore only when there is a perceived better choice); whereas, in the other, it involves random exploration. The performance was analyzed using a reinforcement-learning model. The performance metrics include: 1) accuracies (defined as the proportion of choosing the lowest true shock probability option) and 2) persistent exploitation (defined as the proportion of repeatedly choosing the options with best perceived value).
Results: While there was no significant difference in accuracies between versions (paired t-test, p = 0.1273), persistent exploitation showed a significant difference between versions (paired t-test, p = 0.0005). Additionally, the persistent exploitation difference score showed a strong correlation with the state STAI score (R = -0.55, p = 0.0035).
Conclusions: Participants exhibited higher value-driven exploration as compared to random exploration, evidenced by higher persistent exploitation in the version favoring random exploration. Further, individual differences in persistent exploitation between versions decrease as a function of state anxiety scores. Thus, state anxiety might influence exploratory behavior, with higher scores indicating an ambivalence between random and value-driven exploration.
P18: Two-dimensional Oscillatory neural network model for primary auditory cortex
Abstract:
It has been observed in the auditory cortex of bat and other mammals that they follow a tonotopic architecture, i.e. the neurons in the cortical sheet follow a special architecture. The neurons responding to auditory signals with nearby frequency located close to each other along a spatial dimension. In the previous modelling study, we have proposed a novel coupling strategy which allows multiple supercritical Hopf oscillators to oscillate at a certain “normalized phase difference”. It has been shown that the steady state normalized phase difference among the oscillators can be either of the possible solutions depending on the initial value of the oscillators. Finding the vector space of initial condition for which the system archives the desired solution is not known. The proposed a network of oscillators with unidirectional power coupling connection from a “reference oscillator” solves this issue as well as it has been shown that a 2D array of critical Hopf oscillators receiving unidirectional input from a reference supercritical Hopf oscillator through complex coupling can essentially extract Fourier like representation of input power signal.
P19: Novel feed forward CPG with tunable parameters in a supervised manner
Abstract:
Currently neural networks have been instrumental in solving a variety of sequential processing problems. Complex multi-layer perceptrons have been used for processing complex inputs. We make use of these established paradigms and combine them with Hopf oscillators to create different types of networks that can be used for signal processing. We demonstrate how these networks can be designed to act as signal encoders as well as Central Pattern Generators to control robotic locomotion as well as modelling experimentally recorded EMG signals. In this work we showcase multiple ways to combine layers of real or complex perceptrons as well as MLPs with layers of Hopf oscillators to build different hybrid feed-forward neural network architectures. We propose novel supervised learning rule to train the parameters of the hybrid network to learn the three principle features associated with a signal: frequency, amplitude and phase offset.
P20: Modeling the retention of short term memories by using oscillators
Abstract:
Short term memory helps us to navigate the world on a regular basis but it is limited in capacity as well as duration. The human brain is capable of storing approximately seven short term memories. Storage of these short term memories appear to take place in some neurons by firing continuously after a brief input resulting in neuronal oscillations. Each short term memory is stored in different high frequency oscillations(gamma range) which are nested inside a low frequency oscillation(theta range). These low frequency oscillations repeat, forming a periodic time series signal in which the each short term memories repeat in subcycles. In this work, a complex backpropagation algorithm is combined with an oscillatory network to model this time series signal where approximately seven short term memory patterns are repeated.
P21: Modelling EEG signal recorded during various modes of sleep to analyze the fundamental discriminatory properties
2 Neural Dynamics and Neural Engineering Lab, School of Neuroscience, Virginia Tech., USA.
Abstract:
At the time of sleep EEG recording shifts from low amplitude high frequency to larger amplitude slow frequency. Sleep EEG has four different stages each of having different frequency characteristics i.e. wake, NREM, REM. NREM have four subcategories (1, 2, 3, 4). NREM 3, and NREM 4 are also called slow wave (0.7 to 4.5 Hz). REM sleep occur 20-25% of total sleep (loss of muscle tone happen here). Average power spectrum of NREM is higher than REM. Wake stage is basically alpha activity(10Hz).Sleep spindle occur at NREM(N2) stages(11-15Hz). In adult NREM and REM sleep cycle period repeats in every 90-100Mins. We are proposing a oscillatory reservoir model where a network of Hopf oscillators with tunable natural frequency and complex recurrent connections to model cortical activity during sleep. There are two consecutive phases of training: In the first phase the parameters of the oscillatory reservoir are trained using multiple duration of EEG signal recorded during various phases of sleep whereas in the next phase multichannel EEG signal is modelled using the trained oscillatory reservoir network in the first phase of training.
P22: Making foresighted vs. shortsighted decisions in gambling tasks- untangling the effects of utility vs. Gain-Loss frequency in neuropsychological assessments
NEATLabs, Department of Psychiatry, University of California, San Diego, USA.
Abstract:
Rewards or punishments are contingencies of behavior that guide future decision-making. Decision making based on advantageousness for future such as by maximizing expected values (EV) or utilities of choices have been classically investigated using four choice decks through the Iowa Gambling Task (IGT) in clinical populations as a neuropsychological assessment tool. However many study variants of IGT highlight the imbalance in gain:loss (GL) frequencies across IGT decks and confront their results: they suggest that humans are instead shortsighted, have higher preferences for optimizing GL frequencies than to EV. Here, we investigate a simplified two choice version that allows for more rapid and lifespan assessments and for the first time can untangle the effects of GL frequency and EV response based effects on choices. Precisely, the task has two separate blocks, baseline and experimental, where the baseline has equal expected values and variance of rewards across options and is suitable for assessing the GL preferences, and the experimental block has unequal EVs for the two options whose relative differences to baseline can inform the EV advantageous decision effects. In this study, we present subjective EV effects of about 200 human participants and reliably relate them to clinical measures such as depression, hyperactivity as assessed through their self-reports. We also present reinforcement learning model simulations to estimate subjects risk sensitivity effects on EV based choices. Finally, EEG source correlates for subjective EV guided foresighted decision making is presented.
P23: Event-related spinal responses to omission of expected somatosensory stimulation
1 Department of Behavioral Neurology, Leibniz Institute for Neurobiology, Magdeburg, Germany;
2 Department of Neurology, University Hospital Magdeburg, Magdeburg, Germany;
3 Department of Systems Physiology of Learning, Leibniz Institute for Neurobiology, Magdeburg, Germany.
Abstract:
Predictive coding has become a standard framework to understand perception. Stimulus expectation is known to shape sensory processing at cortical and sub-cortical levels. However, involvement of spinal cord in sensory processing in response to stimulus expectation has not been investigated. We investigate whether stimulus expectation impacts sensory processing at the level of the human spinal cord. To build up stimulus expectation, we applied surface electrical stimulation to the left median nerve at a regular frequency. 15% of the stimuli were randomly omitted to examine time-locked responses to expected, but omitted somatosensory input. Such omission responses at cortical and subcortical levels have previously been taken as signatures of stimulus prediction. Here, we show omission of stimulus related spinal responses occur. The omission of expected stimulus related responses hints towards predictive coding at the level of spinal cord.
P24: Designing subsumption architecture heuristics for adaptive pursuit, patrolling and shielding
Department of Computer Science & Engineering , Indian Institute of Technology Kanpur, Kanpur. India.
Abstract:
In this poster, we present our findings from exploration into low-cost, biological feasible robotic AI. Our objective is to develop an efficient, lightweight control framework for autonomous robotic agents to promote self-sustaining task-driven behavior, and independent coordination with other agents (if present). Our system makes use of a subsumption-based architecture, relevant properties from swarm intelligence, and optimized heuristics to enable speed, scalability, and distributed decision making across multi-robot teams. We propose that this could be a feasible model for tasks that need a high level of adaptability to dynamic elements in a limited environment. Additionally, since the bot's behavior is an emergent property of a bio-inspired control system of asynchronous modules, it would be possible to incrementally test and develop additional functionalities and extend its operation to new domains of deployment if required.
P25: A Comprehensive Rehabilitative Gaming System (cRGS) for stroke rehabilitation of upper extremity
Abstract:
Stroke is one of the leading causes of acquired disability in adults. However, this is often curable with proper physiotherapy. But due to the low therapist to patient ratio, scarce rehabilitation centers, high costs and highly repetitive therapy leading to lack of motivation, most patients do not continue the treatment. Virtual Reality with its engaging aesthetics and rich variety is emerging as a beneficial tool for rehabilitation. However, not many stroke units in India provide this facility. Thus, we propose a 'Comprehensive Rehabilitative Gaming System' (cRGS), which aims at affordable, Stroke Rehabilitation at home, targeting upper extremity. The system consists of Virtual Environments (VE), equipped with multiple games, incorporating physiotherapy. The games are designed such that the complexity gradually increases with the levels thus ensuring player engagement and consequent improvement. To make the experience more interactive, we ensure that the games enable the patients to make movements of exploratory nature.
P26: Goal Pursuit, Voice and Diversity
Department of Management Studies, Indian Institute of Technology Madras, Chennai. India.
Abstract:
We highlight our collective research spanning varied domains of judgment, decision making, diversity and employee voice. As organizations can no longer afford to overlook the underlying processes that make their employees tick, they are increasingly focused on understanding what drives a diverse workforce and their behavior in terms of making decisions and voicing. Owing to the far-reaching organizational benefits of employee voice, we examine the interpersonal drivers of socially embedded, employee centric voice. We explore relationships between implementation intentions, progress monitoring and feedback, multiple goal pursuit and regulatory focus in goal striving. Furthermore, our research examines affective outcomes of diversity, interactions between various types of diversity, workplace disability and neurodiversity inclusion. As researchers, we contribute to the growing body of literature in organizational science and employ experimental techniques with physiological measures for improved understanding of employee behavior.
P27: A Generalized Reinforcement Learning-Based Deep Neural Network Model for Stimulus Recognition, Stimulus to Action Mapping and Selection
Vignayanandam R Muddapu1; Pragathi Priyadharsini Balasubramani2; V Srinivasa Chakravarthy1 & NEATLabs team2
2 NEATLabs, Department of Psychiatry, University of California, San Diego, USA.
Abstract:
A fundamental set of cognitive abilities enable humans to efficiently process goal-relevant information, suppress irrelevant distractions, maintain information in working memory, and act flexibly in different behavioral contexts. Yet, computational and experimental studies of human cognition and their underlying neural mechanisms usually evaluate these cognitive constructs in silos, instead of comprehensively in-tandem within the same individual. Here for the first, we use a scalable, mobile platform that we refer to “BrainE” (short for Brain Engagement), to rapidly assay several essential aspects of cognition simultaneous with wireless electroencephalography (EEG) recordings, and develop a generalized deep neural network model with convolutional neural network component for stimulus recognition and reinforcement learning model for stimulus to action mapping and selection. Using our computational models with BrainE, we assessed six aspects of cognition including (1) selective attention, (2) response inhibition, (3) working memory, (4) reward processing, (5) flanker interference and (6) emotion interference processing, in N>100 human subjects. In this presentation, we show some experimental and model results for selective attention and distractor processing aspects, and understand common vs. distinct neural dynamics underlying these two different cognitive constructs.
P28: A Flip-Flop Based Deep Neural Network Model for Handwriting Generation
Vigneswaran Chandrasekaran; Sandeep Sathyanandan Nair; Vignayanandam R Muddapu & V Srinivasa Chakravarthy
Abstract:
Solving handwriting generation problem gives rich information on understanding motor control and utilizing temporal relationships in the data. The generation model should consider factors like complexities of curved trajectories, pen point movements across the plane, and velocity components in particular direction. The sequential dependency on the data has to be handled using memory-based models with feedback loops and gating mechanisms. In this work, flip flops that are widely used for sequential circuits in digital electronics are used to model handwriting generation. The encouraging performance of the flip flop model shows the potential use case of simple bistable circuits on solving computationally hard problems as well as helps to model intriguing biological processes.
P29: Muscle Synergy Control During Hand Reach Task on Varying Shoulder Configuration
TCS Research and Innovation Lab, Kolkata., India.
Abstract:
Control of human arm in reaching task is a result of complex neural interaction involving central nervous and musculoskeletal system. Activation of group of muscles are planned through synergistic and coordinated recruitment to reach an optimal strategy. A musculoskeletal model of human arm comprising shoulder, elbow and wrist joint have been designed and is used to calculate muscle activation required to perform three specific reaching tasks by changing shoulder configuration. Muscle synergy have been computed on the simulated activation to find a relation between synergy and energy requirement with the change of rotation and elevation of shoulder and its effect on the motion path of the elbow joint. These findings may help to define optimal joint configuration for a planned range of motion during rehabilitation exercises and also in developing neural prosthesis and myoelectric interfaces for efficient arm motion control.
P30: Effect of Constrained and Reinforcement Induced Movement Therapy on Induced Learned Non-use
1 Neuromechanics Lab, Department of Applied Mechanics, Indian Institute of Technology Madras, Chennai. India.
2 Christian Medical College, Vellore, India.
Abstract:
Learned non-use is a behaviorally reinforced condition in patients with neurological diseases like stroke where the affected limb is not used even when it can be used. Learned non-use can be reversed, at least partially, by constraining the less affected limb and forcing patients to use the more affected limb - constraint-induced movement therapy (CIMT). However, CIMT is tedious to implement, frustrating for patients, and has poor retention. Recent studies have shown that arm choice can be altered in the short-term and long-term through subtle manipulation of rewards during reaching movements in healthy and stroke subjects. In this study we introduced a learned non-use-like condition through a visuomotor rotation task and made a direct comparison between the effect of CIMT and reinforcement induced movement therapy(RIMT) on healthy subjects. The results from the study show that both CIMT and RIMT have similar retention.
P31: A test of Mechanical advantage hypothesis at static destinations during upward and downward translation of thumb while holding an object
Neuromechanics Lab, Department of Applied Mechanics, Indian Institute of Technology Madras, Chennai. India.
Abstract:
Grasping an object stable is a common daily life task. Perturbation introduced to the handheld object disturbs its equilibrium. According to the Mechanical advantage hypothesis (MAH), fingers with larger moment arms tend to produce greater normal force than the fingers with the shorter moment arms to re-establish the equilibrium. In the current study, a five-finger instrumented handle was designed with vertical railing on the thumb side of the handle. Thumb sensor was mounted on the slider platform which slides over the railing. The task was to trace the pattern displayed on the monitor by translating the thumb vertically without any tilt. MAH was supported for the downward translations of the thumb while it was not supported for upward translations. Biomechanical constraint of restriction in the range of motion of the Carpometacarpal joint of the thumb played a crucial part in determining the forces exerted by the individual fingers and thumb.
P32: Energy Crisis in Hippocampus Neurons: The Cause or the Effect of Selective Vulnerability
1 Protein Bioinformatics Lab, Department of Biotechnology, Indian Institute of Technology Madras, Chennai, India.
2 Department of Physics, Chuo University, Kasuga, Bunkyo-ku, Tokyo, Japan.
Abstract:
Alzheimer's disease (AD) is one of the most common neurodegenerative disorders that impact more than 47 million people globally and has a steadily increasing mortality rate. The selective vulnerability of hippocampal neurons located in the medial temporal lobe, followed by the disease's progression to higher cortical areas, is the most distinct phenomenon in neuronal cell death. The underlying cause of the selective vulnerability is still unclear. Transcriptome expression and associated pathway analysis of hippocampus neurons help in understanding the mechanism of vulnerability. In the current work, we mapped hippocampus RNA-seq data with genomic/transcriptomic assembly and identified novel variations using genome analysis toolkit. By employing various in silico tools, we predicted the effect of novel variants on transcription factor (TF) binding. We also analyzed changes in transcript expression and built large-scale hippocampus specific co-expression networks. From functional network analysis, we have identified novel variants and TFs associated with dysregulation of blood vessel morphogenesis, energy metabolism, insulin regulation, mitochondrial function, and gliogenesis. This study proposes that restoration of blood-brain-barrier integrity and glial communication may rescue the vulnerable neuron from energy crisis and cell death.
P33: Bio-Inspired Attentional Search for Targets from Road Video
1 Machine Learning group, Continental Automative Components (India) Pvt. Ltd, Benguluru, India.
Abstract:
The poject is a collaboration between CNS Lab, IIT-Madras and Continential Automative Components (India) Pvt. Ltd. The objective of this collaboration is to take inspireation from neuroscience and apply it to solve the problems of automotive domain. The first project under this collaboration aims on designing an architecture for attentional search of targetd from road videos. Impressive results have been achieved in this project.