Research & Education

AI2C Technologies AG is a Swiss ETH Zurich spin-off specializing in computational thinking. The company develops breakthrough technologies in real-time continual learning (RT/CL) and automatic model recalibration, which are crucial for advanced computational thinking. Their products power 'Computational Thinking' machines designed to work alongside humans, enhancing decision-making across various domains. By integrating computing innovation, scientific principles, advanced mathematics, algorithms, and multidisciplinary knowledge, AI2C's mission is to contribute to the advancement of artificial general intelligence (AGI). The team comprises scientists, engineers, and business innovators with expertise in computational science, artificial intelligence, fluid mechanics, and nanotechnology.

harmonic-oscillator-pinn offers an open-source code implementation for a physics-informed neural network (PINN) applied to a harmonic oscillator. This tool serves as a practical example for understanding and experimenting with PINNs, which integrate physical laws into neural network training. It is specifically designed to accompany a blog post by Ben Moseley, providing a hands-on resource for researchers and students interested in scientific machine learning and the application of AI to solve differential equations. The repository includes the necessary code to replicate the experiments and insights discussed in the associated blog post, making it a valuable educational and research asset.

Machine-learning-experiments is an open-source collection of interactive machine learning experiments, designed for educational purposes and hands-on learning. Each experiment features a Jupyter/Colab notebook, allowing users to understand the model training process, alongside a demo page to observe the model's functionality in a browser. The repository covers various machine learning paradigms, including Supervised Learning (Multilayer Perceptron, Convolutional Neural Networks), Unsupervised Learning (Generative Adversarial Networks), and Recurrent Neural Networks. It supports models trained with TensorFlow 2 and Keras, and provides instructions for local setup, dependency management, and model conversion for web deployment using TensorFlow.js. This project serves as a sandbox for exploring different ML approaches, algorithms, and datasets.

Research is a GitHub repository by PaddlePaddle dedicated to novel deep learning research works. It features a comprehensive collection of top conference papers and competition-winning models, covering key areas such as Computer Vision (CV), Natural Language Processing (NLP), Knowledge Graph (KG), and Spatial-Temporal Data-Mining (STDM). The repository offers detailed descriptions, paper links, and implementations for various tasks within these domains, making it a valuable resource for researchers and developers working with PaddlePaddle. It is open-source and freely accessible, encouraging collaboration and advancement in deep learning.

Simple_Reinforcement_Learning is an open-source toolkit designed for the development and testing of reinforcement learning algorithms. It provides a structured environment for implementing various RL techniques, including stateless problems, Markov Decision Processes, dynamic programming, temporal difference algorithms, DynaQ, DQN, policy gradient, Actor-Critic, PPO, DDPG, SAC, imitation learning, offline learning, MPC, MBPO, goal-oriented reinforcement learning, and multi-agent systems. The toolkit is built to run on Python 3.9, PyTorch 1.12.1, and Gym 0.26.2, making it compatible with widely used machine learning libraries and environments. It serves as a valuable resource for researchers and engineers looking to explore and experiment with different reinforcement learning paradigms.

Superbio.ai is positioned as the world's first community-driven AI store specifically for biology. It offers a platform where researchers can access and utilize AI models and computing resources to accelerate their biological research. The service provides different tiers, including a free Basic plan with limited runs and GPU access, and a Boost plan for researchers requiring higher computing needs, offering boosted runs, priority support, batch jobs, and early access to new features. Superbio.ai emphasizes that it does not use user data to improve its models and only deducts successfully finished jobs from user balances, ensuring data privacy and fair usage.

MOVEdot.ai offers AI agents designed to accelerate hardware engineering tasks, particularly in the analysis of test data. The platform helps engineering teams, especially in automotive, motorsports, manufacturing, aerospace, and robotics, to analyze complex data sets, identify anomalies, and make faster decisions. MOVEdot agents can process 100% of data, reducing analysis time by up to 80% and accelerating iteration cycles by 3x. It connects to various data sources like telemetry, sensor logs, and test standards, providing detailed reports and answers to complex questions in plain English. Proven in demanding environments like motorsports, MOVEdot aims to bring this efficiency to all hardware engineering teams.

food-101-keras is an open-source deep learning project hosted on GitHub, designed for food classification using Keras and Tensorflow. It leverages Convolutional Neural Networks (CNNs) to identify 101 different food classes from the Food-101 dataset. The project demonstrates how to fine-tune a pre-trained Google InceptionV3 model, achieving high accuracy in food recognition. It includes detailed steps for data loading, preprocessing, image augmentation, model training, and evaluation. The repository also provides insights into handling large datasets and exporting models for mobile applications, making it a valuable resource for machine learning practitioners and researchers interested in computer vision and food recognition.

Traverse Technologies provides AI-powered solutions for renewable energy developers to optimize wind farm projects. The platform helps increase the Equity IRR of greenfield wind projects by 0.5% through millions of project permutations and direct-to-IRR and LCOE optimization. Key features include large area prospecting, energy optimization, and road optimization. Traverse offers end-to-end services from prospecting to financial close, allowing developers to assess preliminary feasibility, plan measurement campaigns, monitor quality, and optimize layouts, energy, cost, and IRR. The company emphasizes an industry-leading methodology, unlimited revisions, and no variation orders, ensuring alignment with developer needs.

llms-from-scratch-cn is an open-source educational project by Datawhale, designed to help developers and researchers build large language models (LLMs) from the ground up. It offers a comprehensive learning path, emphasizing practical implementation over theoretical concepts. The project focuses on understanding LLM architecture, providing step-by-step tutorials to construct models such as GLM4, Llama3, and RWKV6. It includes detailed code examples, covering encoding, pre-training, and fine-tuning processes, making it accessible for individuals with basic Python and PyTorch knowledge to delve deep into LLM principles.

PROTAC Scientific-Drug Discovery Pro provides comprehensive computational services to accelerate drug lead discovery. The platform utilizes cutting-edge machine learning and informatics tools, alongside in-house multidisciplinary expertise, to streamline the drug discovery process from strategy development through publication. Services include fragment-based lead discovery, molecular docking, protein homology modeling, pharmacophore modeling, and the creation of customized machine learning models for bioactivity prediction. They also offer training programs and workshops, and support for publishing in top-tier journals and patenting. The company aims to reduce costs and save time for researchers, with financial support options for students and referral programs.

Meddenovo Drug Design is a Lyon-based biotech company specializing in AI-powered, nature-inspired cyclic peptide design. Their platform, Mexa, combines large-scale AI-driven generation with physics-based evaluation and refinement to explore millions of candidate peptides per target. This technology allows for de novo drug design even without prior experimental data, accelerating hit identification in drug discovery. Meddenovo focuses on reducing early discovery risk and unlocking peptide-based modalities for difficult targets, particularly in oncology, radiopharmaceuticals, and metabolic diseases. Their approach leverages the unique properties of cyclic peptides, which bridge small molecules and biologics, offering structural rigidity for selective target binding and versatility for complex biological interfaces.

book_DeepLearning_in_PyTorch_Source is an open-source GitHub repository containing the source code for a book titled "Deep Learning Principles and PyTorch Practice." This resource is designed to help users understand deep learning concepts and their practical implementation using the PyTorch framework. It covers a wide range of topics, from introductory PyTorch concepts to advanced applications like generative models, transfer learning, and reinforcement learning. The repository includes code examples for tasks such as text classification, image style transfer, and neural machine translation, making it a valuable learning tool for students and developers looking to gain hands-on experience with deep learning in PyTorch.

Awesome-DeepLearning-500FAQ is a comprehensive open-source resource designed to help individuals understand deep learning concepts through a question-and-answer format. It covers a wide range of topics, including foundational knowledge in probability, linear algebra, machine learning, and deep learning, as well as specialized areas like computer vision, generative adversarial networks, and reinforcement learning. The content is structured into 18 chapters, totaling over 500,000 words, making it a substantial learning aid. Users can access the material in both HTML and PDF formats, with the HTML version offering direct navigation via anchored links for quick access to specific chapters. This resource is ideal for self-study and for those seeking to deepen their understanding of complex AI and machine learning subjects.

how-to-optim-algorithm-in-cuda is a comprehensive open-source repository dedicated to optimizing algorithms using CUDA. It offers a wealth of resources including code implementations for fundamental CUDA operators like reduce, softmax, and elementwise operations, as well as detailed learning notes and blog translations related to GPU and large language models. The project covers advanced topics such as CUTLASS, CuTe DSL, Triton, and PTX ISA, making it an invaluable learning tool for developers aiming to enhance the performance of their CUDA code. It also includes notes on large language model inference/training optimization and GPU/AI system papers.

GPT2-NewsTitle is an open-source project designed for generating Chinese news titles using the GPT-2 model. It provides a comprehensive framework with super detailed Chinese annotations, making it accessible for developers and researchers. The project features a Streamlit page, allowing for easy deployment and visualization of the news title generation without needing Flask+HTML. It also includes a cleaned and organized Chinese abstract dataset, compiled from various sources like Tsinghua News and Sogou News, which is suitable for training and experimentation. The tool supports model training, testing, and deployment, offering a complete workflow for GPT-2 based generation models.

Hong Kong Quantum AI Lab Ltd (HKQAI) is dedicated to advancing scientific research through its AI-based Quantum Simulation Platform. This innovative platform is designed to accelerate the discovery and development of next-generation materials, including organic light-emitting diode (OLED) materials and lithium-ion batteries. Founded by professors from the University of Hong Kong (HKU) and California Institute of Technology (Caltech), HKQAI leverages big data, machine learning, and scientific computing to provide a robust simulation environment. The lab focuses on research directions that include quantum computing applications, advanced material simulations, and AI-driven scientific discovery, offering solutions for complex scientific and engineering challenges.

raster-vision is an open-source Python library and framework designed for deep learning on satellite, aerial, and other large imagery sets, including oblique drone imagery. It offers built-in support for chip classification, object detection, and semantic segmentation, utilizing PyTorch backends. As a library, it provides a comprehensive suite of utilities for handling all aspects of a geospatial deep learning workflow, from reading geo-referenced data and training models to making predictions and writing out results in geo-referenced formats. As a low-code framework, it enables users to configure experiments for machine learning pipelines, including data analysis, chip creation, model training, prediction, evaluation, and deployment bundling. It also supports cloud execution via AWS Batch and AWS Sagemaker.

Nexco Analytics is a Swiss company specializing in Artificial Intelligence, bioinformatics, and data analysis services for the life sciences sector. They offer tailored solutions for academia, pharma, and biotech, focusing on molecular biology and omics data. Key services include bespoke AI solutions, advanced bioinformatics for NGS data (RNA-seq, scRNA-seq, epigenomics, 3D genomics, spatial transcriptomics), and computational structural biology. Nexco Analytics is particularly noted for its expertise in analyzing the "dark genome" to uncover hidden biomarkers and therapeutic targets. Their ONex platform provides a cost-efficient and fast online solution for standard bioinformatics analyses, processing sequencing data within hours. They emphasize a collaborative approach, offering customized data analysis plans, team augmentation, and statistical consulting, with a commitment to using state-of-the-art techniques and providing detailed, publication-ready reports.

UgenTec, now part of Velsera, offers the FastFinder platform to help molecular labs automate sample flow, perform real-time quality control, and streamline result calling. The platform includes modules like FastFinder Workflow for orchestrating sample-to-result processes, FastFinder Analysis for automated assay result reporting, and FastFinder Genotyper which uses AI for genotyping. FastFinder Insights provides smart dashboards for lab operational intelligence, while FastFinder QC offers in-run stats and real-time alerts. The Studio module allows control over assay configuration and SOP automation, ensuring reproducible and documented interpretation rules. It supports various modalities including PCR, Mass Spec, NGS, NAAT/LAMP, Serology, and dPCR, catering to clinical diagnostics, veterinary, pharma, and AgBio labs.

Machine Learning for Science (ML4SCI) is an open-source organization dedicated to integrating modern machine learning techniques with challenging problems across Science, Technology, Engineering, and Mathematics (STEM) fields. The organization fosters collaboration among researchers, students, and developers to advance the application of AI in scientific contexts. ML4SCI actively participates in programs like Google Summer of Code (GSoC), providing opportunities for students to contribute to open-source projects. It also serves as an umbrella organization, welcoming other projects and organizations focused on machine learning for science, and encourages the publication of scientific articles in peer-reviewed journals by its contributors. The initiative aims to push the boundaries of scientific discovery through AI.

SynthText is an open-source tool designed for generating synthetic text images, primarily for use in computer vision research. It enables the creation of extensive datasets of scene-text images, which are crucial for training and evaluating models focused on text localization in natural images. The tool provides scripts for generating samples, including options for visualizing the output, and supports adding new background images with segmentation and depth-maps. It also offers flexibility for generating text in various non-Latin scripts, with several community adaptations available for languages like Chinese, Arabic, Japanese, Korean, Vietnamese, and German. SynthText is a valuable resource for researchers and developers working on text detection and recognition tasks.

Paddington Robotics is dedicated to advancing embodied AI and robotics, working to bridge the divide between digital intelligence and physical interaction. The company focuses on applying AI and robotics solutions to address real-world challenges. Their work involves developing systems that can perceive, reason, and act in physical environments, moving beyond purely digital applications. This approach aims to create intelligent agents capable of performing complex tasks in the physical world, leveraging AI-driven solutions to enhance autonomy and problem-solving capabilities in robotics.

Spleenlab specializes in developing next-generation AI software for autonomous systems across various industries, including aviation, automotive, defense, and robotics. Their solutions are designed to enable safe, scalable, and certifiable automation. For drones, Spleenlab offers capabilities such as GPS-denied navigation, surveillance, collision avoidance, precision landing, and terrain-following. For vehicles, their AI software supports multi-object detection, sensor fusion, freespace detection, occupancy grids, and GPS-denied SLAM. The company's product suite, VISIONAIRY®, includes modules for camera calibration, spatial detection, object detection, navigation, and localization, all compatible with various hardware, chips, and sensors. SPLEENLAB® Suite further provides video stabilization, an operating system, AI runtime, and simulation tools, ensuring efficient performance with minimal computational resources.