Explore Projects

The orbit module provides means to model orbital/trajectory movement of an object in the solar system and includes the calculation of planet positions, perturbation effects on the orbit. It computes the position and velocity of the object, taking into account the forces applied. It also defines a number of generic coordinate systems, including one for each planet and a local one for the object (spacecraft) to be modelled. This proposal aims at providing a generic structure for the orbit module to perform the basic tasks needed. The main milestones for completing this project are to: Create object {spacecraft} Create other objects {celestial bodies} Define local coordinate system of object {spacecraft} Define coordinate systems for other objects {planets} Model trajectory of an object {spacecraft in solar system} Define perturbation effects on orbit {position, velocity of object, forces applied}

What I aim for the “5G NR module benchmark analysis for distinct channel models” project is to complete it in the available time (175 hours) and successfully, achieve all goals. Moreover, I intend to provide a simple interface, using the NR Helper class, to make setting the channel model easy for users. Afterward, I will create an example of an outdoor scenario to evaluate, assess, and benchmark the channel models that are being implemented. Also, to provide a test case code that uses these three channel models.

In this project, we aim to port DeepChem Models from Tensorflow to Pytorch. I want to select Chemnet Models model to work with. I want to put the Smiles2Vec model as my first priority.

Every time a researcher wants to train a language model, they end up rewriting the same infrastructure: data loading, optimizer setup, training loops, logging, and checkpointing. PyTorch-Ignite solves the engine abstraction, but researchers still have to manually wire everything together for each new experiment. Trainite is a pip-installable Python toolbox built on top of PyTorch-Ignite that handles this. A single trainite init command gives researchers a complete, runnable training project. They bring their model and dataset, and Trainite handles the rest. Deliverables include an installable package, a CLI scaffolding tool, a working end-to-end training pipeline with a built-in Transformer example, YAML-based configuration, organized timestamped outputs, full documentation, and tests.

Scrum Helper currently supports only GitHub and a partially-integrated GitLab, with all platform logic tightly coupled in a 2,190-line monolithic file. This makes adding new SCM platforms nearly impossible without duplicating the entire data-fetching flow. This project refactors Scrum Helper's core into an extensible Provider Pattern architecture by defining a standardized SCMProvider interface that decouples platform-specific API calls from the report generation engine. I will then implement native providers for GitLab (including self-hosted instances), Gitea, and Bitbucket, each returning normalized data through a unified schema. Deliverables: (1) SCMProvider interface with normalized data models, (2) Refactored GitHubProvider with full regression coverage, (3) GitLabProvider with self-hosted baseUrl support, (4) GiteaProvider tested against Codeberg, (5) BitbucketProvider with OAuth 2.0 auth, (6) unified caching and cross-provider error handling, and (7) architecture guide and contributor docs for adding future providers. All 4 platforms will generate identical scrum reports by project end.

This project aims to extend OpenCV’s DNN module with efficient runtime support for Large Language Models (LLMs), focusing on enabling Visual Language Models.

Machine learning algorithms are heavily relied on to understand the data generated at the European Council for Nuclear Research's (CERN) Large Hadron Collider (LHC) which produces immense amounts of data through the measurement of the products of particle collisions. Since data from these events can typically be formed into a graph structure, deep geometric methods, such as graph neural networks (GNNs), have been used as an approach to various HEP anlysis tasks. One such task is jet tagging, where jets are viewed as point clouds with distinct features and edge connections between their constituent particles. To enhance the validity and robustness of deep networks, embedded symmetries, which are present in many physically realizable datasets and fundamental physical theories, can also be discovered by the model through the use of invariant I(n) and equivariant networks E(n). However, due to the increasing size and complexity of these datasets, as well as the models which process them, an involved yet computationally inexpensive approach must be utilized. To handle an increase in the complexity of the data, I suggest the implementation of a classical attention (AT) transformer with the necessary modifications to incorporate, invariant or equivariant, and GNN layers. As the extension of classical algorithms via bit-wise processes to quantum algorithms via qubit-wise processes has been rapidly developing over the last few years, I propose the full development of general quantum invariant and equivariant graph attention transformer (QIGAT or Quantum Informer, QEGAT or Quantum Equiformer) architectures to meet the demands of requiring these machine learning algorithms to offer enhanced robustness through the adherence to fundamental physical principles while reducing their computational cost.

This project would implement the Besag, York, Mollie (BYM) model in PyMC. BYM is a highly flexible model for studying spatial data and is used widely in epidemiology, agriculture, and ecology. The strategy behind the BYM model is to treat the outcome of interest as the result of three factors: some set of predictor variables, spatial covariance between neighboring regions, and random effects that represent non-spatial heterogeneity. Although very flexible, the BYM model can be difficult to specify in a way that is simultaneously computationally efficient, interpretable, and identifiable. Recently, Morris et al (2019) demonstrated an alternative specification of the BYM model that is significantly more efficient, interpretable, and can be identified with Monte Carlo Markov Chain (MCMC) samplers. Developing an implementation in Python with PyMC would make the model more accessible to a greater variety of users. Furthermore, BYM models are only available in a Bayesian framework. As a leading Bayesian statistics package, PyMC should support this extremely useful model in spatial statistics.

<p>Details of deep learning models and their performance are, unfortunately, often published without accompanying implementation code. Those models that come from speech recognition seem to be particularly susceptible to this phenomenon. The present project proposes to create a working implementation of a speech recognition model using the Flux library for the Julia programming language and contribute its code to the Flux model zoo. The model to be implemented is Zhang et al.'s (2017) model from their paper "Towards end-to-end speech recognition with deep convolutional neural networks." Due to being implemented using only convolutional layers, this model will be lighter to train than previous models that have used heavier recurrent layers, while still achieving state-of-the-art performance. Having a working implementation of this network will be a step forward in opening the culture of automatic speech recognition. As a result, newcomers to the field will have a recent example to look at for inspiration, which is paramount because there are not many novice-friendly resources available for doing speech recognition research.</p>

<p>Deepchem is one of the premier open-source deep learning libraries for chemistry. It is used by people worldwide, thus there is a need to develop and provide access to updated cutting-edge models, more metrics and tests for model performance evaluation and more recently developed loss functions. My GSoC ‘21 Project proposal is to add the Molecule Attention Transformer and a protein structure model to Deepchem and/or Jaxchem.</p>

Improve security model and processes for Kyverno. Document security processes, help define a threat model with risks and mitigation and add best practice processes like publishing signed images.

This project aims to implement the sequence space jacobian methods of Auclert et al. (2021) into the Heterogeneous Agents Resources and toolKit (HARK) under the econ-ARK project. These methods would allow the HARK toolkit to provide open source software for economic researchers to solve heterogeneous agent macroeconomic models. To implement these computational methods, objects and tools in HARK will be extended to allow for the efficient simulation of a macroeconomic model, the computation of Jacobians to linearize a model, and the computation of impulse responses to a linearized model. At the completion of this project, economic researchers will be able to simply specify the equations that define their model and be immediately returned the solutions to their model.

Gemma Facet is a comprehensive platform that provides an end-to-end solution for fine-tuning Gemma language models through a microservices architecture. The platform implements four core services: dataset preprocessing with support for local and Hugging Face datasets, automated training jobs using Unsloth and Transformers libraries, model inference capabilities, and flexible export functionality supporting multiple formats (adapters, merged models, and GGUF). The backend leverages Google Cloud Run services for scalable compute, Firestore for database operations, and Google Cloud Storage for artifact management. The frontend is built with Next.js, Tailwind CSS, and Shadcn UI, providing an intuitive dashboard interface. Infrastructure is managed through Terraform with containerized services using Docker. The system handles complex workflows including dataset splitting configurations, asynchronous training job management, and multi-format model export pipelines. Key technical implementations include Cloud Run Job integration for resource-intensive operations, comprehensive API design with full documentation, and optimized data processing pipelines for efficient model fine-tuning workflows.

This project ports the Chemception model—a deep CNN for molecular property prediction—from TensorFlow to PyTorch.DeepChem has been shifting its backend from TensorFlow to PyTorch. This transition requires the porting of models that were previously built in TensorFlow to PyTorch.

This project aims to improve the reward signal used in the RL-based training of MLGO models. I plan to integrate an existing GNN-based performance estimation model into the training process to generate said reward signal, with some key improvements to facilitate better modelling of the execution environment and runtime related behaviors, such as branch mispredictions and cache misses. An outline of those improvements would be to create a corpus of blocks with a variety of enhancements over the original (first deliverable), modifying the model to be compatible with the new dataset (second deliverable), and experiment with feeding the model additional context beyond what is available within individual basic blocks to allow the model to factor in aforementioned runtime behaviors to gain valuable insights on where to go forward.

The aim of this project is to focus on patch-level classification tasks within the domain of whole slide tissue images. The approach involves utilizing pre-trained models as encoders and applying them to the task of patch-level classification. By leveraging pre-trained models, readily available on platforms like Hugging Face, aim to expedite the development process while maintaining high performance. The ultimate objective of the project is to integrate these pre-trained models into caMicroscope, providing users with a convenient tool for downloading and training models tailored to their specific tasks. Deliverables: - A documented method for training a classification model using a pre-trained or foundation model. - Implementation of this method as a set of software components. - Integration of the components with caMicroscope.

SOFIE aims to streamline the integration of deep learning models into the ROOT scientific software framework through efficient inference code generation. By implementing missing ONNX operators, particularly those used in transformer models, we will enhance TMVA's capabilities for building complex neural networks. Our project will deliver modular operator classes adhering to ONNX standards, accompanied by thorough unit testing, enabling seamless deployment of deep learning models in physics software and analysis frameworks.

<p>In recent years, deep learning has made remarkable progress in the area of NLP and CV with big data. In the area of chemistry, Graph Convolutional Networks, in which compounds and crystals are regarded as graphs, have attracted much attention.Deep learning frameworks for molecules such as DeepChem, Chainer Chemistry, and Deep Graph Library have emerged to more easily exploit these models of deep learning. On the other hand, these frameworks did not support inorganic crystal and the model support was biased towards Graph Convolutional Networks.JAX Chem is a JAX-based deep learning library for complex and diverse chemical modeling. JAX Chem’s motivation is to support not only models of Graph Convolutional Networks, but also Attention-based models like Transformer, and even models for wave functions involving complex physical equations like SchNetOrb or FermiNet. JAX is suitable for such complex modeling because JAX’s API allows you to write models declaratively. In the period of GSoC, I will implement the standard GNN model, the dataset utilities and the GCN model for MaterialInformatics, and the Attention-based model, as the beginning of JAX Chem.</p>

According to the Google Summer of Code (GSoC) project ideas listed for One proposed DeepChem project involves adding support for the OLMo Large Language Model (LLM) family to the DeepChem repository using the Hugging Face Model wrapper. The goal of this project is to enable DeepChem users to perform tasks such as text generation, classification, and continued pretraining with large language models, including Open Language Model (OLMo). I have already contributed several improvements to DeepChem’s HuggingFaceModel integration, including improving typing and validation for the fill mask method and implementing a new generate() function that wraps HuggingFace’s text generation API. I have also already demonstrated partial compatibility with OLMo models through recent contributions to the generate() function and example scripts using allenai/OLMo-1B-hf. I have also added support for causal language modeling in the HuggingFaceModel wrapper, including integration with AutoModelForCausalLM and end-to-end validation through unit tests. My contributions demonstrate familiarity with the DeepChem codebase and provide the foundation for extending DeepChem to support large language models such as OLMo.

<p>A differential equation model of a dynamical system is a nonlinear parameterized model that is created to match realistic scenarios and occasionally it might be associated with some data, obtained from the system or through simulation. Once there is reasonable level of confidence in the correctness of the model, the task that remains is to estimate the parameters of the model. Due to the structure of parameter estimation problems in dynamical models, statistics and machine learning techniques are an ideal choice for determining the parameters. During the course of the summer I will be implementing some statistical algorithms, including Stochastic Approximation Expectation Maximization(SAEM) and Maximum A Posteriori Estimation (MAP), for parameter estimation of a dynamic model. I will also work on extending support for parameter estimation in Stochastic Differential Equations (SDEs) by adding first differences distribution to generalized Log-Likelihood. These would be quite important additions to the suite of methods in JuliaDiffEq and would be of great use to scientific community involved in systems biology, HIV-AIDS study, and drug dosage estimation.</p>

<p>In order to make DBpedia and its humongous linked data available to a larger user base in their natural languages, a Neural QA model has been developed to answer the question in English posed by users. This particular project aims to make our end-to-end system learn better compositionality of questions, by improving our current dataset and our learning model.</p>

The project incorporates addition of a new video filter to FFmpeg, PU21 transform. PU21 has been developed to convert absolute high dynamic range (HDR) linear color values into approximately perceptually uniform (PU) values, which can be used with standard quality metrics. This transformation would eventually help evaluate HDR videos based on quality tests known for SDR ones.

The project aims at building foundational models suitable for characterizing the vast astronomical data and emphasizing their use case for various downstream tasks. Collecting suitable datasets, training the models on simulated datasets and applying them on real observational datasets. The main purpose of the project is to develop a suitable deep learning approach for these astronomical datasets which is able to overcome the specific instrument differences for the different observations.