Research & Education

TrafficFlowPrediction is an open-source project designed for predicting traffic flow using various neural network architectures, including Stacked Autoencoders (SAEs), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU). This tool is ideal for researchers and data scientists working in transportation planning and traffic management. It requires Python 3.6, Tensorflow-gpu 1.5.0, Keras 2.1.3, and scikit-learn 0.19. Users can train models with their own data, with experiment data from the Caltrans Performance Measurement System (PeMS) provided as an example. The project offers detailed metrics like MAE, MSE, RMSE, MAPE, R2, and Explained variance score for each model, demonstrating its effectiveness in traffic forecasting.

ZynqNet is an open-source project stemming from a Master Thesis, focusing on FPGA-accelerated embedded convolutional neural networks. It provides a comprehensive solution for image classification on embedded systems, featuring the ZynqNet CNN, an optimized and customized CNN topology, and the ZynqNet FPGA Accelerator, an FPGA-based architecture for its evaluation. The project also includes the Netscope CNN Analyzer, a custom tool for visualizing, analyzing, and editing CNN topologies. ZynqNet is designed for high efficiency, achieving 84.5% top-5 accuracy with minimal computational complexity, making it ideal for real-time and power-constrained applications. The repository offers the full project report, CNN prototxt, pretrained weights, HLS C++ source code for the accelerator, and firmware for the Zynq XC-7Z045 ARM processors.

temperature_scaling is an open-source Python module designed to calibrate neural networks by adjusting their confidence scores. Originally created as a demonstration for PyTorch 0.3, it implements temperature scaling, a post-processing technique that divides logits by a learned scalar parameter to minimize negative log-likelihood on a validation set. This helps address the common issue of neural networks outputting overconfident probabilities, ensuring that confidence scores better match true correctness likelihood. While the repository is unmaintained, it offers a clear example of how to integrate temperature scaling into a project for improved model calibration.

Applying_EANNs is a 2D Unity simulation designed to showcase how cars can learn to navigate various courses. The cars are controlled by a feedforward neural network, whose weights are optimized using a modified genetic algorithm. This project provides a practical demonstration of evolutionary artificial neural networks in a simulated environment. Users can tinker with simulation parameters in the Unity Editor or run the built executable with default settings. The neural network architecture includes an input layer, two hidden layers, and an output layer, with its training managed by a customizable genetic algorithm. The user interface displays real-time data for the best performing car, including neural network output, evaluation value, and a generation counter, along with a visual representation of the neural network's weights.

MiniWorld is a minimalistic 3D interior environment simulator specifically designed for reinforcement learning and robotics research. It allows users to simulate environments featuring rooms, doors, hallways, and various objects, making it suitable for tasks like training AI agents in office, home, or maze-like settings. Written 100% in Python, MiniWorld is easily modifiable and extensible, offering features such as few dependencies, good performance, lightweight design, and support for domain randomization for sim-to-real transfer. It also provides fully observable top-down views, depth map production, and the ability to display alphanumeric strings on walls. This project has been deprecated as of August 11, 2025, and is no longer receiving updates or support.

DIPY (Diffusion Imaging in Python) is a comprehensive open-source Python library designed for the analysis of MR diffusion imaging and other 3D/4D+ medical images. It provides a robust set of generic methods for tasks such as spatial normalization, signal processing, machine learning, and statistical analysis. Beyond general medical image processing, DIPY specializes in computational anatomy, offering advanced techniques for diffusion, perfusion, and structural imaging. The library is intended for research purposes, with a clear disclaimer for clinical deployment. It supports installation via pip or conda and adheres to Scientific Python SPEC 0 for version compatibility, making it accessible for researchers and developers in the medical imaging field.

Domain Transfer Network (DTN) is a TensorFlow-based implementation for unsupervised cross-domain image generation. This tool enables users to transfer image characteristics from one domain to another, such as converting SVHN images to MNIST, without requiring paired training data. It is designed for researchers and developers interested in image synthesis and domain adaptation, providing a practical framework for experimenting with generative models. The repository includes Python scripts for dataset download, preprocessing, model pretraining, training, and evaluation, making it a comprehensive resource for those working with generative adversarial networks (GANs) and similar architectures.

MatterSim is a deep learning atomistic model developed by Microsoft, designed for simulating materials across a wide range of elements, temperatures, and pressures. It enables researchers and scientists to predict and analyze material behavior using advanced deep learning techniques. The tool offers two pre-trained models, MatterSim-v1.0.0-1M and MatterSim-v1.0.0-5M, based on the M3GNet architecture, with the larger version providing higher accuracy. Users can install MatterSim via PyPI or from source, and it supports finetuning on custom datasets. While primarily for bulk materials, it can be fine-tuned for specific applications like surfaces or interfaces.

OmniIsaacGymEnvs offers a robust platform for developing and testing reinforcement learning agents within the Omniverse Isaac Gym ecosystem. It leverages PPO from the rl_games library and is built upon Isaac Sim's omni.isaac.core and omni.isaac.gym frameworks. Users can train policies, load pre-trained models for inference, and run simulations in both graphical and headless modes for optimized performance. The tool supports various tasks, from Cartpole to complex robotic simulations like Humanoid and ShadowHand, and provides extensive configuration options via Hydra. It also integrates with Docker for streamlined deployment and offers livestreaming capabilities for real-time visualization.

Model Medicines is an AI-driven company dedicated to building better medicines by innovating at the intersection of data science, biology, and drug development. The platform utilizes AI to model chemistry and human biology, accelerating the discovery and development of life-changing drugs. With 192 compounds and 67 validated assets in disease-relevant cellular models across 12 therapeutic targets, Model Medicines focuses on areas such as virology, oncology, inflammation, and longevity. Their proprietary GALILEO™ and AmesNet™ technologies enable ultra-large virtual screening and agentic AI breakthroughs, leading to the identification of best-in-class potential therapeutics, such as MDL-001, a direct-acting, broad-spectrum antiviral.

PINNpapers is a comprehensive, open-source repository maintained by the IDRL lab, dedicated to curating essential research papers on Physics-Informed Neural Networks (PINNs). Since PINNs have gained significant traction in scientific computing, this resource serves as a valuable collection of representative works in the field. The repository categorizes papers across various aspects of PINNs, including foundational models, parallel computing approaches, acceleration techniques, model transfer and meta-learning, probabilistic PINNs, uncertainty quantification, and diverse applications. It also lists relevant software libraries like DeepXDE and SciANN, providing links to papers and code where available. Researchers and practitioners can use this resource to stay updated on the latest advancements and foundational concepts in PINN research.

SnakeFusion is an AI project that leverages genetic algorithms and neural networks to train virtual snakes within a game environment. The core concept involves training five individual snakes, which can then be fused together to create a single, more advanced 'ultimate snake'. This project serves as a practical demonstration of applying evolutionary algorithms and AI in game development. Built using Processing, it provides a hands-on approach to understanding how AI can learn and adapt. Users can interact with the system by adjusting mutation rates, saving trained snakes, and initiating the fusion process to observe the creation of a 'super snake'.

squeezeDet is an open-source project providing a TensorFlow implementation of SqueezeDet, a convolutional neural network specifically designed for real-time object detection. This tool is particularly optimized for autonomous driving applications, emphasizing a unified, small, and low-power architecture. It allows users to train and evaluate object detection models using datasets like KITTI, supporting various network backbones such as SqueezeNet, ResNet50, and VGG16. The repository includes scripts for installation, demo execution, training, and validation, making it a comprehensive resource for researchers and developers working on efficient object detection in resource-constrained environments.

spark-py-notebooks is a comprehensive collection of IPython/Jupyter notebooks designed to educate users on various Apache Spark concepts using Python (pySpark). The tutorials range from fundamental to advanced topics, focusing on Big Data Analysis and Machine Learning. Users can learn about RDD creation, basic RDD operations like map, filter, and collect, sampling, set operations, and data aggregations. The collection also delves into working with key/value pair RDDs and introduces MLlib for basic statistics, exploratory data analysis, logistic regression, and decision trees. Additionally, it covers Spark SQL for structured processing with DataFrames and includes applications like building a movie recommendation web service.

Setup-NVIDIA-GPU-for-Deep-Learning is a comprehensive, open-source guide designed to assist users in setting up their NVIDIA GPUs for deep learning tasks. It outlines a clear, step-by-step process, starting with the installation of the latest NVIDIA GPU drivers. The guide then proceeds to cover essential software components such as Visual Studio with C++ support, Anaconda/Miniconda for package management, the CUDA Toolkit, and cuDNN. Finally, it provides instructions for installing PyTorch and includes a script to test the GPU setup, ensuring all components are correctly configured for optimal deep learning performance. This resource is invaluable for deep learning practitioners and AI researchers looking to streamline their development environment setup.

tf-gnn-samples is a GitHub repository offering TensorFlow implementations of various Graph Neural Network (GNN) architectures. It serves as the code release for an article introducing GNNs with feature-wise linear modulation (GNN-FiLM). The repository includes implementations for Gated Graph Neural Networks (GGNN), Relational Graph Convolutional Networks (RGCN), Relational Graph Attention Networks (RGAT), Relational Graph Isomorphism Networks (RGIN), GNN-Edge-MLP, and Relational Graph Dynamic Convolution Networks (RGDCN). It provides scripts for training and evaluating models on tasks such as citation networks (Cora, Pubmed, Citeseer), protein-protein interaction (PPI), quantum chemistry prediction (QM9), and variable misuse detection (VarMisuse). The code allows users to reproduce experimental results presented in the accompanying research paper, making it a valuable resource for researchers and developers working with GNNs.

Abzu is a biotechnology company leveraging explainable AI to innovate in the field of RNA therapeutics. The company specializes in developing best-in-class RNA drugs, including siRNAs, ASOs, and anti-miRs, for significant medical needs. Their AI-guided design platform, powered by the QLattice®, allows for the in silico exploration and prioritization of vast sequence spaces, evaluating over 100,000 design variants to predict efficacy and developability properties. This approach significantly reduces experimental cycles, lowers costs, and shortens the time to candidate selection. Abzu also focuses on RNA-based delivery systems, developing targeted aptamers for cell-specific uptake of therapeutic RNA, offering a modular platform for precision delivery beyond the liver. The team combines deep RNA biology, AI-driven design, and drug development experience to create a closed learning loop where data refines models and models improve molecules.

Additive Catchments is dedicated to restoring river health by providing advanced infrastructure for water quality monitoring. The platform utilizes sensor networks to deliver real-time data, offering transparent insights and actionable intelligence crucial for effective water management. It aims to give rivers a voice by integrating environmental data, civic infrastructure, and pollution monitoring to create a comprehensive river health index. This tool is designed to support water governance and catchment management, enabling stakeholders to make informed decisions and build a sustainable future where rivers, communities, and society can thrive.

vec2text is an open-source library providing utilities for decoding deep representations, such as sentence embeddings, back into text. It enables users to train various architectures that reconstruct text sequences from embeddings and also run pre-trained models. The library supports both direct inversion from embeddings and inversion of text strings, with options to refine results through multiple steps and increased search space. It is particularly useful for researchers and developers working with text embeddings and language models, offering functionalities like interpolation of embeddings and detailed guidance on training custom inversion and corrector models.

ttt-rl is a reinforcement learning example implemented in C, designed to teach the basics of reinforcement learning through a tic-tac-toe game. The neural network learns to play against a random adversary from scratch, without any pre-existing knowledge of the game. It uses a simple architecture with a single hidden layer and is contained in under 400 lines of C code, with no external libraries. This project is particularly valuable for programmers, especially young programmers, who want to understand new fields through small, self-contained, and well-commented C programs. It demonstrates how RL can learn complex behaviors from basic reward signals.

TorchDrug is a robust, PyTorch-based machine learning platform specifically designed for drug discovery. It simplifies the implementation of graph operations in a PyTorchic style with GPU support, making it accessible even for practitioners with minimal drug discovery knowledge. The platform facilitates rapid prototyping of machine learning research by providing a wide range of common datasets and building blocks. Users can easily work with graph-structured data and molecules, extracting properties without deep domain expertise. TorchDrug also accelerates training and inference across multiple CPUs or GPUs, offering seamless scalability for complex experiments. It supports integration with Weights & Biases for experiment tracking and management.

Brightseed is a continuous innovation platform designed for health and life sciences companies. It unifies deep science, cutting-edge AI, and the world’s largest proprietary bioactive dataset to streamline the discovery, validation, and commercialization of new products. The platform helps teams identify and prioritize promising opportunities earlier, reducing guesswork and increasing confidence before significant investments are made. By replacing fragmented R&D workflows with a unified system, Brightseed accelerates progress from discovery to decision, leading to faster cycles, quicker time to market, and a higher probability of success for innovations. It empowers teams to deliver more successful innovations, faster, fueling sustainable growth and long-term differentiation.

X2Paddle is a deep learning model conversion tool developed under the PaddlePaddle ecosystem, designed to help users of other deep learning frameworks quickly migrate their models and projects to PaddlePaddle. It supports the conversion of prediction models from major frameworks like Caffe, TensorFlow, ONNX, and PyTorch. Additionally, X2Paddle facilitates the migration of entire PyTorch training projects, including both training and prediction code, to the PaddlePaddle framework. The tool offers detailed API comparison documentation to reduce the time and effort developers spend on migrating models. It boasts support for a wide range of models, covering over 130 PyTorch OPs, 90 ONNX OPs, 90 TensorFlow OPs, and 30 Caffe OPs, making it a comprehensive solution for model migration.

Cypher AI offers intelligent digital infrastructure tailored for modern biology and life science R&D. It addresses the need for real-time adaptable systems by generating experimental designs, managing sample tracking, and integrating instruments. The platform also orchestrates analysis pipelines and CRO (Contract Research Organization) activities on demand, streamlining complex research processes. This allows scientists to accelerate discovery and innovation by providing a comprehensive solution for managing and executing biological research, from initial design to final analysis.