Explore Projects

<p>Modeling the volatility of financial markets is central in risk management. A seminal contribution in this field was the development of the GARCH model by Bollerslev (1986) where the volatility is a function of past asset returns. The GARCH model is today a widespread tool in risk management. However, recent studies show that estimates of GARCH models can be biased by structural breaks in the volatility dynamics (Bauwens et al., 2010; Bauwens et al., 2014). These structural breaks typically occur during periods of financial turmoil. Estimating a GARCH model on data displaying a structural break yields a non-stationary estimated model and implies poor risk predictions. A way to cope with this problem is provided by Markov-switching GARCH models (MSGARCH) whose parameters vary over time according to some regimes. These models can quickly adapt to variations in the unconditional volatility level, which improves risk predictions (see Ardia, 2008).</p>

<p>Re-building the official Tensorflow models to make it TF 2.0 compatible. This project proposes holistic improvements to the models repository to upgrade models/research and models/official. The project scope also includes building new deep learning models and features to improve research prototyping with Tensorflow 2.0. Creating model migration guides for the official models will enable onboarding to TF 2.0 with eager mode for R&amp;D and graph mode for serving.</p> <p>The official models (including Mask R-CNN, StackGAN, Age-cGAN) will be upgraded with tf.data pipelines and distributed training with DistributionStrategies. Other improvements include bug fixes and the use of tf.GradientTape to compute the gradients more efficiently using Autodifferentiation. The Autoencoders and the GANs projects will be recreated with TF 2.0 with features to train efficiently and export to deploy.</p>

<p>We aim to construct a probabilistic model which explains a hypothetical criminal's behavior based on that criminal's actions. We will be given the criminal's motions and the crimes they have committed, and from this, we will try to find a probabilistic model that explains their behavior. We are planning to construct a Hidden Markov Model, which is a probabilistic system describing the transitions between the criminal's actions.</p> <p>Treating the criminal's decisions about what locations to go to, and which crimes to commit, as the result of a Markov process, we can use statistical methods to reconstruct a model that predicts their behavior. For this project will use a simulated criminal whose actions are generated by an agent-based model (resulting from a previous GSoC project). We should be able to create a model that is nearly equivalent to the original program. With the reconstructed model, the program should be able to make predictions about the criminal's future actions.</p>

Particle colliders such as the Large Hadron Collider (LHC) play a crucial role in advancing our understanding of fundamental particles and their interactions. To achieve their desired performance, these experiments rely on accurate simulations of high-energy physics (HEP) events. However, simulating HEP events required for these experiments with traditional approaches such as Monte Carlo simulation is time-consuming and computationally expensive because of which there is a need to explore different approaches. DeepFalcon is a generative ultra-fast non-parametric detector simulation package. This project aims to extend DeepFalcon by adding functionality for Diffusion models to it. The proposed project will build upon existing research on diffusion models and leverage the advancements in deep learning techniques to explore the application of diffusion models for fast simulation of HEP events in particle detectors and potentially improve the accuracy and efficiency of the simulations.

<p>In the upcoming summers, I aim to achieve six objectives in three different fields. These are:</p> <h3>1. Model Import :</h3> <p>I plan on making a reader for machine learning models in Julia. This will help us in loading and running models made using different frameworks into Julia. This involves making a reader for ONNX models, and Keras models. I collaborated with the mentor, Mike Innes to work on this and the code can be found in the ONNX.jl repository under FluxML organization.</p> <h3>2. Model Export :</h3> <p>I will work on demos for the FluxJS.jl model zoo. This will make Machine Learning in Julia more dynamic and flexible, as it would help us in visualizing models in the browser. I'll also work on the main FluxJS package, as it still needs a lot of code tracing.</p> <h3>3. Computer Vision :</h3> <p>I propose to work on two different projects in this direction. The first project involves integrating Metalhead.jl with ONNX.jl . This will help in expanding the functionalities offered by Metalhead. The second part would include loading and running as many ONNX models as possible.</p>

Optical flow is the problem of estimating the motion of objects in an image or video sequence. Optical flow is pivotal to many computer vision applications such as object tracking, video stabilization, and motion analysis, providing essential information about the dynamics of a scene or an object. With the upsurge in the use of deep neural networks, DNN models were developed to solve the optical flow problem, achieving state-of-the-art results by learning to estimate motion from image data. However, DNNs can become computationally expensive making it infeasible to be deployed on embedded systems. This problem is tackled by the development of lightweight DNN models. However, OpenCV model zoo hasn't had any implementation of an optical flow lightweight model yet. Thus, this project aims to find the best lightweight optical flow model in terms of model size, speed, and accuracy to introduce to the OpenCV model zoo.

Online learning has taken the front seat in the post pandemic age. With the advent of sophisticated AI architectures like the Transformers, it is but natural that AI would find its way into education. Learning online via YouTube or MOOC platforms is often done as a method of self learning. The biggest obstacle faced by a student in self learning is the lack of attention span. An online tool that can generate short quizzes on input educational content can be of great use to teachers and students alike as it can help retain important information, frame questions and quickly revise large chunks of content. EduAid is one such project where I propose the usage of open source and lightweight Large Language Models for quiz generation. The language modelling capabilities of LLMs can be leveraged for achieving this task. Finally, we develop a chrome extension that allows end users to interact with the pipeline and generate quizzes based on input text.

Current deep learning models for CMS particle reconstruction are often task-specific. This project aims to develop versatile foundation models by pre-training deep neural networks on unlabeled CMS End-to-End (E2E) detector data using various Self-Supervised Learning (SSL) strategies (MAE, I-JEPA, CAE, VAE). These models learn robust, general-purpose representations of particle interactions. We will then comprehensively benchmark these foundation models by adapting them for diverse downstream tasks including Particle Identification (Classification), Kinematics Estimation (Regression), Detector Super-Resolution, and High-Fidelity Event Generation (investigating VAEs and conditional Diffusion Models). Deliverables include the open-source codebase with reproducible pipelines, pre-trained model weights, detailed benchmarking results/analysis, and user-friendly documentation.

Facility location modeling, a set of geoanalytic methods, is a critical problem solving tool leveraged in public and private sector decision making and planning activities. These models can be applied to a number of real-world problems, providing solutions that can account for varied resource availability and dispersion needs. Presently, PySAL has a basic implementation of location models, but these tools require further enhancement to become a comprehensive resource for users. The proposed 350-hour project aims to implement functionality for users to calculate results for the Backup Coverage Location Problem, optimize resources using the p-Dispersion model, and add additional functionality to existing models, such as facility capacity, demand unit shape, and distance metrics. Functionality will be implemented using well-known open-source GIS packages such as networkX, geopandas, and shapely.

The KServe Models Web App is a crucial front-end interface for managing ML models deployed through KServe, a core component of Kubeflow. However, the application has become outdated, with deprecated dependencies and technical debt. This project aims to modernize and enhance the app by upgrading its tech stack, resolving open issues, and implementing a robust CI/CD pipeline with GitHub Actions. Key improvements include regression testing with Cypress, model editing capabilities, and synchronization with KServe 0.14+. Additionally, the project explores migrating the app from kserve/kserve-models-web to kubeflow/kserve-model-ui for better integration. The deliverables include a stabilized, modernized codebase, improved user experience for ML engineers, and streamlined contributor documentation. This effort will reduce operational overhead, enhance usability, and support Kubeflow’s mission of making ML workflows simpler and more scalable.

Current LHCb calorimeter reconstruction uses rule-based clustering without exploiting timing or learned representations. The PicoCal upgrade will provide picosecond timing and longitudinal segmentation, creating an opportunity for ML-based reconstruction. I propose a space-time kernel transformer that embeds timing as a physics-informed spatial coordinate for energy reconstruction, benchmarked against the Cellular Automaton and Graph Clustering baselines across four incremental dataset stages. Deliverables include a complete end-to-end pipeline, trained models with timing ablation study, benchmark tables with physics and latency metrics, and a documented open-source PyTorch Geometric codebase for the LHCb group.

This project aims to implement the Fourier Neural Operator (FNO) in DeepChem, expanding its support for solving complex partial differential equations (PDEs) in scientific machine learning. Traditional neural networks often struggle with PDEs due to their reliance on fixed-grid discretization. FNO addresses this by operating in the frequency domain, learning resolution-invariant mappings between function spaces using the Fourier transform, making it both scalable and efficient. The project involves developing a new SpectralConv layer and a high-level FourierNeuralOperator class compatible with DeepChem’s PyTorch backend. The implementation will be modular, enabling future extensions such as Physics-Informed Neural Operators (PINO). Deliverables include the FNO model, unit tests, and a comprehensive tutorial.

This project develops linear-scale attention vision transformers for CMS End-to-End jet classification and proxy mass regression on 8-channel 99.1% sparse detector images. Five architectures were trained and benchmarked under identical conditions during the test task: XCiT with domain MAE pretraining (AUC 0.9237, 2.42x S/sqrt(B)), XCiT with ImageNet pretraining, XCiT from scratch, a custom L2ViT combining local window and linear global attention built from scratch, and Swin-T as the softmax baseline. All models use a dual head for joint classification and proxy mass regression. 11 analysis studies were completed beyond standard benchmarking: channel importance (Ch0 and Ch3 critical, AUC collapses to 0.47 without either), channel pair interactions, Grad-CAM saliency, t-SNE feature representations, layer-by-layer CKA similarity, data efficiency (5x labelled data savings), jet multiplicity analysis, energy threshold study (detector implications for zero-suppression), S/sqrt(B) significance curves, adversarial robustness, and calibration with temperature scaling. Key findings: domain MAE pretraining outperforms scratch by +0.125 AUC and leads to fundamentally different representations (CKA 0.008-0.015 at every layer), linear attention beats softmax by +0.063 AUC under identical conditions, and block 1 of the pretrained model alone exceeds the fully trained scratch model in feature separation. The summer work extends this to multi-class classification (quark/gluon/top/W/Z) with per-class channel importance ablations, token pruning exploiting 99.1% sparsity, sparse convolution baselines, hybrid CNN+ViT architectures, and comparison against ParticleNet and JEDI-net on the same held-out test set.

This project aims to advance the SYMBA framework for calculating squared scattering amplitudes in high-energy physics by incorporating physics-informed machine learning principles. While recent transformer-based approaches have shown strong performance by framing amplitude squaring as a sequence-to-sequence problem, they largely rely on syntactic patterns and do not explicitly capture the physical structure of Feynman diagrams or the semantic roles of symbolic entities. This work proposes to bridge that gap by integrating structural and physics-aware representations into the learning process. By embedding domain knowledge into model design, this project seeks to improve accuracy, generalization, and physical consistency, particularly for complex processes such as QCD where symbolic expressions grow rapidly in complexity. The outcome will contribute towards more scalable and interpretable machine learning approaches for symbolic physics computations, while providing reproducible tools and insights that can benefit the broader ML4Sci and high-energy physics communities.

<p>The following modules are to be integrated with caMicroscope:</p> <ul> <li>Creating a workflow to allow model developers to allow their model to be run on a selected image or region of interest by the client, to identify cellular features or cancer.</li> <li>Add the latest deep learning models researched in the area of digital pathology to caMicroscope to help the pathologists.</li> <li>Improve the segmentation app.</li> <li>A tutorial to enable users to make models compatible with caMicroscope.</li> </ul> <p>The motto of this project is to make the recent research in deep learning more accessible</p>

Monoclonal antibodies (mAbs) are a potent therapeutic because of their ability to precisely target molecular surfaces at a sub-protein level resolution. They are used in cancer checkpoint blockades, treatments for viral infection, and useful in situations of snakebites as anti-venom. Current methods of antibody discovery involve inoculating various chimeric animal constructs with specific antigens and isolating and purifying antigen-specific molecules. As such there has been a significant interest in designing antibodies in-silico. A growing number of large language models have been employed to capture the sequence distributions of these proteins, with varying degrees of success. However, the vast majority of language models are encoder-only transformers whose target-conditioned sequence generation capabilities are limited, and their code is prototypical rather than production ready. This project aims to integrate and finetune a target-conditioned LLM for eptiope-specific antibody design. Extending DeepChem's codebase to include support for this class of therapeutics would not only help democratize this technology for resource constrained operations, but also bring these groups onto a platform with a lively user base and healthy collaborative environment.

<p>The Toolkit for Multivariate Data Analysis with ROOT (TMVA) provides a machine learning environment for the processing and evaluation of multivariate classification, both binary and multi-class, and regression techniques targeting applications in high-energy physics. The latest development in TMVA is a fast inference system, with its own intermediate representation of the deep learning model compliant with the ONNX standard. With this new inference system, TMVA aims to provide convenience for users in High Energy Physics in the deployment of their Deep Learning models in a production environment. To facilitate the usage, storage and exchange of these models,this project aims to develop the storage functionality of Deep Learning models in the ROOT format, popular in the High Energy Physics community.</p>

<p>This proposal aims to describe a set of transformations to implement for the spirv-fuzz tool. It provides a high-level overview of the transformations as well as specific implementation details that might be important to mention.</p>

Real-time object tracking models working on edge devices. I will use MAE pre train and knowledge distillation to make the performance better. The deliverables include a high-speed model and the c++ code to deploy it

<p>The Machine Learning team at the Wikimedia Foundation works with the aim of building ethical ML solutions to help Wikimedia communities and teams in the pursuit of open knowledge. In this pursuit, there is now a need for a new technical infrastructure to help scale these solutions even further. The current system, ORES has been enabling ML at the foundation for ~6 years. The proposed infrastructure, Lift Wing aims to enable more widespread participation and collaboration with volunteers and communities by lowering the bar to contribute models to the system.</p> <p>The goal of this project will be to train ML models that perform equivalent or better than the existing models without using revscoring and the ORES infrastructure.The goal is to achieve equivalent or better performance on the same task given the original data used to train the models and ensuring that all the libraries they use are open-source and industry standard (e.g tensorflow, pytorch, scikit learn, hugging face, etc) with the exception of mwparserfromhell which is required to parse the wikitext that is included in the data.</p>

<p>In this project we will look at loop transformation heuristics, such as the unroll factor.</p>

LLM hallucinations have been tackled using various techniques, including modern retrieval methods such as Retrieval-Augmented Generation (RAGs). These techniques show that combining non-parametric memory, which is retrieved and ranked from sources like knowledge graphs and/or databases, with the LLMs' parametric memory helps generate less factually incorrect information. However, hallucination explainability still remains unsolved, especially on logical reasoning tasks. Additionally, the impact of uncertainty caused by these hallucinations on human trust and decision-making still remains unexplained. To address these problems, we propose two solutions: 1. Conduct a first-of-its-kind study that develops a new conversational explanation paradigm under compositional uncertainty and performs human evaluation. 2. In parallel, develop a neuro-symbolic framework for explaining hallucinations in large language models by utilizing agentic Prolog knowledge-base rules. To accomplish these objectives, we first build a lightweight full-stack web application that collects quantitative and qualitative human feedback for model alignment. We also design LLM explanation paradigms grounded simultaneously in multiple sources (open and closed source), develop models (LLMs and agents), and build a pipeline for storing the human feedback in common data format to a secure database, which can be used to further adjust the weights of the reward model.

Alaskan coastlines are rapidly evolving due to the direct impacts of climate change. Rising sea levels, melting sea ice, and shifting environmental conditions make accurate modeling of coastal changes a pressing need. Current approaches often miss key environmental variables like tidal shifts, temperature variation, and wave energy. Moreover, manual coastline digitization remains time-consuming and inefficient, limiting scalability. The aim of the project is to develop and deliver a highly accurate, automated model that can extract vectorized coastlines directly from PlanetLabs satellite imagery. This will build upon the existing pipeline and model by re-training the water detection (NDWI-based) and segmentation methods to perform well on PlanetLabs’ imagery. Further aims include: Ensuring the extraction is robust in challenging conditions such as water shadows, buildings, artifacts, and especially steep, cliff-like coastal areas. Integrating additional data sources if needed, such as SWIR data the project will also explore the use of auxiliary data sources (Landsat & sentinel) and elevation data for addition of topographic information, to improve accuracy. Ultimately producing a finished model that can process large amounts of imagery and accurately predict coastline changes over time, which is essential for improving coastal erosion forecasts in Alaska considering the effects of climate change.