Explore Projects

<p>This project is about building an <em>embodied cognitive</em> simulation, i.e. that in which robots we call <em>vehicles</em> have a body and a simple "mind", represented by a neural activational network. The body has a defined shape, activator sensors that capture signals from the environment, and motors, that move the vehicle as a reaction to the signals. We then evolve neural networks inside vehicles using a Genetic Algorithm with an appropriate fitness function, and hope to observe some natural behavioural patterns as well as certain connectome motifs seen in nature. This would allow us to reproduce the very same behaviours’ simulation, as well as hypothesize on correspondence of connectome motifs to specific behaviours. This project's results have possible applications in brain development studies, as well as transferring synthesized behaviour models to robots and enriching virtual embodied systems’ intelligence (e.g. game AI).</p>

The main objective of this project is to generate a 3D model based on a set of 2D axolotl embryo images to aid as a research tool. The problem can be approached by solving 3 sub categories namely : 1) Extracting embryo images from given set of 2D images using Region Of Interest (ROI) extraction techniques. 2) Generating a 3D model by extracting spatial information based on contours of the embryo in different orientations. 3) Projecting the extracted embryo images on the generated 3D model to get a finished 3D model.

This project centers on the enhancement and integration of the CellSAM and DevoNet models to elevate the analysis of 3D microscopy images stored in TIFF format. CellSAM, initially a segment-anything model, was fine-tuned to allow promptable segmentation of cellular structures in microscopy images. Conversely, DevoNet, a dual vision encoder model, operates seamlessly on 3D TIFF files. One encoder is trained on segmentation maps while the other specializes in analyzing the centroid position of cells. This synchronized operation aids in the precise calculation of cellular volume, area, and centroid locations. The collaborative mechanism between the fine-tuned CellSAM and DevoNet models provides a robust framework for more accurate and insightful analysis of 3D cellular microscopy images, thereby broadening the understanding of cellular morphologies and interactions within a three-dimensional space.

NeuroML and PyNN are open, simulation independent formats that are used to convert large scale network models. Network models incorporating realistic connectivity such as network structure, cell or synapse properties are being widely used to understand information processing in cortical structures. Moreover, making the models available on Open source brain repository along with documentation will enable the computational neuroscientists community to test them across multiple simulator implementations. The major goals of the project are - 1. Verify, improve and test the original model code in NetPyNE locally 2. Convert the published network model involving channels, cells, connectivity developed in a simulator specific format to NeuroML 3. Document and illustrate the baseline expected behaviour of the model and sharing them across Open Source Brain repositories

<p>The Virtual Brain (TVB) is a neuroinformatics platform for the simulation of the dynamics of large-scale brain networks with biologically realistic connectivity. However, the fMRI data modeling performed by TVB suffers from the drawback of considering a standard model for the underlying Hemodynamic Response Function (HRF), throughout a brain and across various brain models. Several studies have indicated that HRF shows variability across both the subjects and across various brain regions of a particular subject. To account for the same, we aim at extending the rsHRF-toolbox to interface between estimating region-specific resting-state HRF (rsHRF) from fMRI data, and modeling BOLD activity in TVB. The extended toolbox shall take fMRI data as input, estimate the region-specific rsHRF for regions (network nodes) defined by the connectome used in TVB, and mediate the fMRI data simulations through TVB's Bold monitor and BoldRegionROI monitor while accounting for the underlying HRF variability across the network. We also aim to build a GUI interface for the toolbox and consequently a docker implementation for the same as we expect end-users from various communities and backgrounds.</p>

This project will convert Macaque auditory thalamocortical circuits model based on NetPyNE implementation into NeuroML standard formats and testing it across multiple simulation engines to ensure that it produce the same results. I'll convert the whole model.

This proposal outlines a novel computational framework for modeling neural developmental programs using Growing Hypergraph Neural Networks (Growing HNNs). These models serve as unified representations by integrating both spatial and lineage dynamics during the embryo-genetic development. This is achieved by encoding cells, spatial structures, lineage relationships, and developmental cues into a dynamic hypergraph, this approach enables end-to-end simulation and prediction of neural growth, differentiation, and functional emergence. The framework is biologically inspired, scalable, and aligned with recent advances in graph neural networks and developmental biology. The goal is to model the cross-connections between diverging networks (anastomoses) using Hypergraphs during the embryo-genetic development.

ASSR refers to the cortical entrainment to frequency and phase of an auditory signal that is presented in a fixed “train of clicks”, in a gamma range rhythm (40 Hz). A hallmark of schizophrenia is a reduction in ASSR; this project aims at reproducing this phenomenon using an auditory cortex (A1) model with thalamocortical connectivity. The latter simulates a cortical column with a depth of 2000 μm and 200 μm diameter, containing over 12k neurons and 30M synapses. Specifically, we aim at reproducing results from an experiment looking at the effects of increased CB1 receptor availability and GABA receptor deficits, as these have been linked to the EEG abnormalities that characterize schizophrenia. We will start by running batch simulation tasks to pull out connectivity rules from the A1 model, modifying GABA and CB1 Receptors. This step will be run on a scaled version of the A1, so that I will be able to use a personal computer, but I may employ cloud resources as well.Then, we will analyze parameter sweeps for local field potentials, using the LFPy toolbox. Afterwards, we will move to parameter optimization of the A1 model so that it reproduces the ASSR, using the Optuna HPO toolkit. Our ultimate goal is to reproduce the ASSR phenomenon in the A1 model. For every step of the process, documentation will be made available on a deployed site.

The application of Network Control Theory to medical neuroimaging consistently faces the approximation crisis. Current tools rely on infinite-horizon continuous models, binary structural masking and statistically inaccurate harmonization. They fail to capture the discrete, finite time and stochastic truth of brain dynamics grounded in biology. This limits their use in computational psychiatry and biomarker discovery. NeuroSim is intended as a Python model built from scratch to resolve these problems. It aims to simulate in-silico brain stimulation with rigorous physics-based constraints. Over the 350 hour period, the three prime modules will be made available to the community:- 1. Discrete Finite Horizon Physics: A transition from standard iterative summations to discrete time Van Loan Doubling algorithm, significantly reducing computational load for large scale neural networks. 2. GraphNet based Laplacian Regularization: Implementation of a proximal gradient descent module to generate soft-prior connectivity matrices. This resolves the problem of structural blindness in binary DTI-based masking. 3. Bias-less Harmonization and Ground-truthing: Deployment of a NeuroCombat protocol to ensure data security in multi-site studies. Validation to be done against non-linear Wilson-Cowan neural mass simulations. Using the above modules, NeuroSim positions itself as a highly scalable, diagnostic-adjacent toolset capable of modelling complex neural mass transitions. It establishes a standard for next generation non invasive neuromodulation and macro-scale brain stimulation.

As the project title describes the main goal is to create a model class that builds, saves, loads, predict and fit user created models. This can be achieved by saving the metadata of model with model configuration and the model data itself and save it in a particular file format. Secondary goal of the project is to work on updation of the docker container of PyMC and write tutorials for deployment of PyMC models (using above API calls) for most common deployment tools such as docker, sagemaker, ML-Flow airflow and dask.

<p>Unlike traditional inverse identification tools that rely on gradient and gradient-free methods, simulation-based inference has been established as the powerful alternative approach that yields twofold improvement over such methods. Firstly, it does not only result in a single set of optimal parameters, rather simulation-based inference acts as if the actual statistical inference is performed and provides an estimate of the full posterior distribution over parameters. Secondly, it exploits prior system knowledge sparsely, using only the most important features to identify mechanistic models which are consistent with the measured data. The aim of the project is to support the simulation-based inference in the brian2modelfitting toolbox by linking it to the <code>sbi</code>, <code>PyTorch</code> powered library for simulation-based inference, development of which is coordinated at the Macke lab.</p>

<p>The Calabrese Lab 8-cell Leech Tutorial that is described by Hill et al 2001 has been a staple for teaching computational neuroscience at Emory University for many years, and it has also been used in various summer courses. This tutorial is not only a fully constructed 8-cell circuit that can generate heart rhythms, but also a great teaching tool thanks to the visual interface where a student can turn on synaptic connections or change maximal ionic conductances. However, the original tutorial has been developed using the now obsolete Genesis simulator. Unfortunately, running the Genesis simulator nowadays requires complicated software set up that prevents many non-technical students from accessing the tutorial. We had previously started porting the tutorial to a more modern format that can be executed through a web browser (see 8-cell Leech Heartbeat Network Model Tutorial), increasing the accessibility of this classic tutorial for teaching and research purposes. At this summer’s GSoC, we would like to finish this port and make the tutorial available. We selected the Neuron simulator language for running the model using NeuroML and Python as the description languages.</p>

<p>ViSP supports CAD model in .cao format or in .vrml format which are needed for Markerless model-based tracker module. Currently, the creation of .cao model formats, which is a homemade file format, and its corresponding config files has to be done manually. The goal of this project is to automate this process by providing extensive tools to achieve perfect and loss-less conversion from existing 3D file formats to .cao and also generation from scratch:</p> <p>1) Develop dedicated Blender plugin to edit and convert from classical 3D file format (for example .obj) to home-made CAD model file format.</p> <p>2) Develop a very simple CAD model viewer and editor that allows to add ViSP MBT specific characteristics to a face or a line such as a label or a level of details (lod).</p>

This project studies the development of a tiny worm called Caenorhabditis elegans, focusing on how its cells grow and interact from its early stages. It uses two special tools called Spatial Hypergraphs and Lineage Hypergraphs to map out and understand these interactions. Spatial Hypergraphs look at how cells are arranged and interact in space, which helps us understand how they communicate and influence each other. Lineage Hypergraphs track the family tree of each cell, showing how cells evolve and form different parts of the worm's body. By combining these tools, the project aims to get a complete picture of the worm's development, from a single cell to a fully formed organism. This work could help us learn more about basic life processes and improve medical and tissue engineering fields. The project is supported by international organizations interested in neuroscience and the use of advanced computer models to study brain and nervous system development.

<p>The Openworm foundation is building the world's first digital organism.(a worm matrix!) The DevoWorm project(under Openworm) is building a physics-based simulation of mosaic embryogenesis, with application to the nematode Caenorhabditis elegans. This initiative will focus on incorporating secondary data from nematodes into an XML-based computational framework. The model-building will result in an XML specification of embryo physics that describes mosaic developmental process. This specification will be used to build trees and networks that describe relationships between individual cells. This will provide the Openworm Foundation and the larger neuroscience research community with an informatics framework for understanding neural precursor cells and developing nervous systems. It will pave the way to unlock the mysteries of the evolution of a nematode’s nervous system from a single cell to the fully developed nervous system in that of an adult nematode (in this case, the 302 neurons of C.elegans).</p>

The proposed project aims to address the sustainability challenges faced by open-source software projects by using an agent-based modelling and simulation approach. Open-source projects often face issues such as limited resources, difficulty in attracting and retaining contributors, and communication breakdowns, which can hinder their growth and sustainability. To address these challenges, the project will simulate various scenarios and identify the factors that contribute to the success or failure of open-source projects. The project will provide a framework that enables simulating different scenarios to assist project maintainers in making informed decisions that promote sustainability. This approach can offer a more detailed and context-specific understanding of the challenges open-source projects face and develop effective strategies for maintaining their sustainability.

DevoGraph is an open-source framework that models cell developmental process as graphs, extracts and converts latent information into high-level embedding using tools of Graph Neural Networks. The embeddings can be utilized in a series of downstream tasks regarding C. elegans from the perspective of network, such as analyzing cell tracking, cell division and differentiation, lineage trees, cell network structures in embryogenesis, etc.

<p>Many symbolic program analysis techniques use satisfiability modulo theory (SMT) solvers to verify properties of programs. SMT solvers can provide solutions quicker if they cache their results. GREEN, which is currently integrated into Symbolic Pathfinding (SPF), is a promising SMT memoization solution. The project entails (a) optimization of GREEN's satisfiability checking, (b) improving the model counting and (c) incorporating unit propagation into GREEN.</p>

<p>In this project, we seek to develop 3D meshing and visualization techniques that are better adapted to the complexity of vascular structures.</p>

My project proposal for GSoC in Ivy is to work on building Vision Models using Ivy and creating demos and tutorials to showcase the implementation of these models. These models include PSMNet and MLP-Mixer.

<p>Adding support for archive storage (i.e in the form of .zip or .tar.* family) of models for all pre-existing models in DFFML and update tests, documentation and fix any model specific bugs that come along the way while implementing this feature. This will not only save the model state but also all the configuration of a model. There would be two benefits of this implementation to the user:</p> <ol> <li>Increased Reproducibility of DFFML models. </li> <li>Better Portability of DFFML models.</li> </ol>

<p>DFFML provides APIs for dataset generation and storage, and model definition using any machine learning framework, from high level down to low level use is supported. As the goal of DFFML is to build a community driven library of plugins for dataset generation and model definition, so that developers and researchers easily plug and play various pieces of data with various model implementations or generate datasets using the implemented features to increase the accuracy of output. For this, DFFML needs to implement large number of machine learning models as well as various features. I have planned to add the below listed Models/Algorithms to DFFML.</p> <ol> <li>Model 1: Ordinary Least Square Regression (OLSR)</li> <li>Model 2: Logistic Regression</li> <li>Model 3: k-Nearest Neighbour (kNN)</li> <li>Model 4: Naive Bayes</li> </ol>

This project proposes the integration of Empirical Dynamic Modeling (EDM) into the Fisheries Integrated Modeling System (FIMS), an open-source framework used for fisheries stock assessment. While FIMS currently supports parametric models such as catch-at-age and surplus production models, ecological systems often exhibit nonlinear dynamics that are difficult to capture with predefined equations. EDM offers a data-driven alternative that reconstructs system dynamics directly from time-series observations using delay embedding techniques. The project will implement core EDM components within FIMS, including delay-embedding generation and prediction algorithms such as Simplex projection, S-map, and Gaussian Process EDM. These methods will be integrated with the existing FIMS statistical inference framework to enable empirical forecasting and hybrid modeling approaches. The final outcome will be a modular EDM module, tested using GoogleTest and testthat, along with documentation and example workflows to demonstrate its application in fisheries modeling.