Research & Education

foldingdiff is an open-source research tool developed by Microsoft that utilizes diffusion models to generate novel protein backbone structures. It employs trigonometry and attention mechanisms, as detailed in its preprint on arXiv. The tool provides a trained model on HuggingFace spaces and SuperBio, allowing users to generate protein structures directly from a browser. It supports installation via conda and pip, and offers scripts for training custom models, downloading necessary data, and sampling protein backbones. Additionally, foldingdiff includes functionalities for evaluating designability through inverse folding with ProteinMPNN or ESM-IF1, and structural prediction using OmegaFold or AlphaFold2.

Machine-learning-for-proteins is a comprehensive and collaborative listing of academic papers focused on the application of machine learning in protein research. This resource aims to keep pace with the rapidly evolving field of protein engineering and analysis, offering a broader scope beyond engineering-specific applications. Papers are categorized by application and model type, such as reviews, tools and datasets, generative models, and various prediction tasks (stability, structure, sequence, interactions). Within each category, papers are listed in reverse chronological order, providing the most up-to-date information. It encourages community contributions through pull requests or issues, making it a dynamic and continuously updated resource for anyone interested in the intersection of machine learning and protein science.

MiniMax-M1 is the world's first open-weight, large-scale hybrid-attention reasoning model, powered by a hybrid Mixture-of-Experts (MoE) architecture combined with a lightning attention mechanism. Developed based on the MiniMax-Text-01 model, it features 456 billion parameters with 45.9 billion parameters activated per token. A key differentiator is its native support for a context length of 1 million tokens, significantly larger than competitors. The lightning attention mechanism ensures efficient scaling of test-time compute, consuming 25% of the FLOPs compared to DeepSeek R1 at a generation length of 100K tokens. MiniMax-M1 is trained using large-scale reinforcement learning (RL) on diverse problems, including mathematical reasoning and software engineering. It offers two versions with 40K and 80K thinking budgets, outperforming other strong open-weight models on complex software engineering, tool-using, and long-context tasks. It also supports function calling capabilities and provides a chatbot and API for general use and evaluation.

Matterport3D is a comprehensive open-source dataset designed for RGB-D machine learning tasks. It includes data captured from 90 properties using a Matterport Pro Camera, offering a rich resource for researchers and developers. The repository provides raw data, derived data, annotated data, and scripts/models for various scene understanding tasks such as image keypoint matching, view overlap prediction, surface normal estimation, region type classification, and semantic voxel labeling. It also includes tools for loading and viewing the data, making it a valuable asset for advancing research in indoor environment understanding.

ml-pen-and-paper-exercises is a comprehensive, open-source collection of pen-and-paper exercises designed to deepen understanding in machine learning. Each exercise is accompanied by a detailed solution, making it an excellent resource for self-study or as supplementary material for courses. The collection covers a wide array of topics, including linear algebra, optimization, directed and undirected graphical models, expressive power of graphical models, factor graphs and message passing, inference for hidden Markov models, model-based learning (including ICA and unnormalised models), sampling and Monte-Carlo integration, and variational inference. The exercises are compiled into a PDF available on arXiv, and the project is licensed under a Creative Commons Attribution 4.0 International License.

motion_imitation is a code repository accompanying the paper "Learning Agile Robotic Locomotion Skills by Imitating Animals." It provides a Gym environment for training a simulated quadruped robot to imitate various reference motions, offering example training code for learning policies. The tool supports Python 3.7 or 3.8 on Ubuntu, MacOS, and Windows, and can be installed as a pip package. It includes features for training and testing imitation models, working with motion capture data, and implementing locomotion using Model Predictive Control (MPC). The repository also details how to run MPC on real A1 robots, making it a comprehensive resource for researchers and developers in robotic locomotion.

neurodiffeq is an open-source Python library built on PyTorch, designed for solving ordinary and partial differential equations (ODEs and PDEs) using neural networks. It provides a flexible framework for implementing existing techniques of using artificial neural networks (ANNs) to approximate solutions. Unlike traditional numerical methods, neurodiffeq aims to compute continuous and differentiable solutions. The library supports various features including solving systems of ODEs and PDEs, handling initial and boundary conditions, and customizing network architectures. It also offers tools for monitoring training progress, implementing transfer learning, and defining custom sampling strategies for training points. Additionally, neurodiffeq supports solving solution bundles and inverse problems, making it suitable for complex scientific and engineering applications.

Nuvo AI is a human-centered research lab dedicated to architecting Africa's cognitive infrastructure from first principles. The organization focuses on building advanced reasoning systems for the future. While specific product details are not extensively outlined, the core mission revolves around developing AI that understands real-world contexts and solves meaningful problems in collaboration with people. This initiative aims to amplify human judgment, creativity, and decision-making through innovative AI solutions, contributing to the advancement of AI capabilities within the African continent.

Neural-Network-Visualisation offers an interactive web-based visualization for a compact multi-layer perceptron, specifically trained on the MNIST handwritten digit dataset. Users can draw digits on a 28x28 grid and observe in real-time how activations propagate through the 3D network. The tool also displays prediction probabilities, providing a clear understanding of the neural network's decision-making process. It highlights the strongest incoming connections per neuron and uses color-coding to represent activation sign and magnitude. The project is open-source and provides Python helper scripts for training the MLP and exporting weights, supporting Apple Metal, CUDA, or CPU acceleration. It also includes a timeline export feature, allowing users to scrub through different training checkpoints.

oie-resources offers a comprehensive, curated list of resources focused on Open Information Extraction (OIE). This GitHub repository serves as a central hub for researchers and academics, providing access to a wide array of materials including research papers sorted chronologically and by category, code implementations, and datasets. It covers not only core OIE systems but also related work such as taxonomizing open relations and various downstream applications like Question Answering, Knowledge Base Population, and Event Extraction. The resource also features information on OIE systems for different languages, supervised OIE, PhD theses, and demos, making it an invaluable reference for anyone working in the field of natural language processing and information extraction.

The Open Persian ASR Leaderboard is a platform designed for evaluating and ranking Automatic Speech Recognition (ASR) models specifically for the Persian language. It enables users to submit their own ASR models by providing the model name in the format "user_name/model_name" and have them assessed against a standardized benchmark. This tool facilitates comparison of different models, helping researchers and developers identify top-performing ASR systems for Persian. The leaderboard provides a transparent and accessible way to track advancements and performance metrics in Persian ASR, fostering competition and innovation within the field.

Qvantify automates qualitative research by using AI to conduct remote interviews, allowing organizations to gather hundreds of deep feedback interviews daily. The platform enables continuous discovery, market expansion assessment, and solution validation by letting AI interact with customers and users. It combines the simplicity of surveys with the depth of in-depth interviews, offering features like white-labeling for brand customization and localization for conducting interviews in over a hundred languages. Qvantify also provides a REST API for data access and future features include a customer portal, voice input, and outbound calling, making it a comprehensive tool for scalable qualitative feedback.

RecommenderSystem-Paper is an open-source GitHub repository that serves as a curated collection of significant papers, tools, and frameworks within the domain of recommender systems. It is designed to assist researchers and academics by providing a centralized resource for reading and exploring key advancements. The repository categorizes papers by conference (e.g., KDD, ICDM, AAAI, WWW, NIPS, ICML, CIKM, SIGIR, Recsys, WSDM) and by interesting topics such as Cold Start and Deep Learning. Beyond academic papers, it also lists useful recommender system engines like Mosaic and Crab, and algorithm frameworks such as Surprise and LightFM, making it a comprehensive resource for understanding and implementing recommender technologies.

relational-networks is an open-source Pytorch implementation of the "A simple neural network module for relational reasoning" paper, also known as Relational Networks. This tool is designed for researchers and developers working on visual reasoning and relational AI tasks. It has been thoroughly tested on the Sort-of-CLEVR task, a simplified version of CLEVR, which involves processing images with various colored shapes and answering both relational and non-relational questions. The implementation demonstrates superior performance compared to traditional CNN + MLP models, particularly in relational reasoning tasks, and includes modifications for improved computational efficiency.

rl-agents is an open-source project providing a comprehensive collection of Reinforcement Learning agent implementations. This tool is designed for researchers and developers working in the field of AI, offering a variety of planning and learning algorithms. It serves as a valuable resource for experimentation and building new RL applications. The project's open-source nature fosters community contributions and allows for flexible integration into diverse research and development environments, making it suitable for both academic and practical applications in reinforcement learning.

slime is an advanced post-training framework designed for Reinforcement Learning (RL) scaling, specifically tailored for large language models. It achieves high-performance training by seamlessly integrating Megatron with SGLang, enabling efficient and scalable operations. The framework supports flexible data generation through custom data workflows, allowing users to adapt to various training requirements. slime facilitates efficient training across different modes, making it a versatile solution for developers and researchers working with large language models and RL applications. Its focus on performance and flexibility makes it suitable for complex AI development tasks.

The Center for Responsible AI is an organization dedicated to fostering the development of ethical and impactful AI products. It brings together a consortium of startups, research centers, unicorns, law firms, and industry leaders to create a virtuous circle of innovation. The center's core mission revolves around three pillars: fair and transparent AI that contributes to a more equal society, sustainable AI that uses energy efficiently, and trustworthy AI that is explainable and complements human capabilities. It aims to develop next-generation AI products based on these principles, with a significant focus on the Portuguese ecosystem. The Center for Responsible AI emphasizes practical solutions and collaboration across various sectors to advance responsible AI.

VideoLLaMA3 is an open-source project offering a series of multimodal foundation models designed for advanced image and video understanding. It provides models like VideoLLaMA3-7B and VideoLLaMA3-2B, which are capable of tasks ranging from general image and video comprehension to more specialized applications such as multi-image comparison, visual referring, and grounding. The project includes detailed instructions for inference, training, and evaluation, making it suitable for researchers and developers. It supports various benchmarks for performance assessment and offers a flexible framework for preparing custom training data. The models are available on Hugging Face, facilitating easy access and integration into AI development workflows.

SRSWTI is a Knowledge and Inference Platform designed to facilitate the understanding and application of knowledge. While specific features are not detailed on the publicly available pages, the platform's core offering revolves around providing tools and resources for knowledge management and inference. It aims to support users in various contexts, likely including educational and research environments, by enabling them to process, organize, and derive insights from information. The platform's focus on "Knowledge and Inference" suggests capabilities related to data analysis, pattern recognition, and potentially predictive modeling, catering to those who need to manage and leverage complex information effectively.

3D-Reconstruction-with-Deep-Learning-Methods is a comprehensive GitHub repository that curates a list of open-source projects centered around 3D reconstruction utilizing deep learning techniques. This resource is specifically designed for AI researchers and deep learning engineers who are actively involved in or interested in the field of 3D vision. The repository provides direct links to various projects, often including details on the deep learning frameworks used (e.g., TensorFlow, PyTorch), associated academic papers, and licensing information. It acts as a central hub for discovering, exploring, and potentially contributing to cutting-edge advancements in 3D reconstruction, making it easier for professionals to find relevant tools and methodologies for their research and development.

Data Phoenix is a comprehensive live media platform designed for AI and Data professionals. It offers a rich array of content including news, in-depth articles, digests, event listings, videos, and original shows. The platform focuses on covering technologies under the hood, best practices, and live demonstrations from industry builders. Users can stay updated on the latest advancements in AI and data, explore tech talks, demo stages, and research talks. Data Phoenix also provides opportunities to subscribe to newsletters, follow live streams, and partner with them to reach a large audience of AI and data professionals.

AIRS, or Artificial Intelligence Research for Science, is an open-source initiative offering a comprehensive collection of software tools, datasets, and benchmarks. It is specifically designed to support research in AI for quantum mechanics, density functional theory, small molecules, protein science, materials science, molecular interactions, biological science, partial differential equations, and ordinary differential equations. The project's goal is to foster an integrated, open, reproducible, and sustainable set of resources to advance the emerging field of AI for Science. It includes various methods and resources, with the list continuously expanding as research progresses, making it a valuable resource for academic and scientific communities.

TinyLLaVA_Factory is an open-source modular codebase designed for building small-scale large multimodal models (LMMs). Implemented in PyTorch and HuggingFace, it emphasizes simplicity, extensibility, and reproducibility. The framework allows users to customize their own LMMs with reduced coding effort and fewer mistakes. It integrates a suite of cutting-edge models and methods, supporting LLMs like OpenELM, TinyLlama, StableLM, Qwen, Gemma, and Phi. For vision towers, it includes CLIP, SigLIP, Dino, and combinations thereof. Connectors such as MLP, Qformer, and Resampler are also supported, alongside various training recipes including Frozen/Fully/Partially tuning and LoRA/QLoRA tuning. The project provides trained models, performance benchmarks, and local demo options for Gradio web and CLI inference.

AttentionDeepMIL offers a PyTorch implementation of Attention-based Deep Multiple Instance Learning, a technique detailed in the paper "Attention-based Deep Multiple Instance Learning" by Ilse, Tomczak, and Welling. This open-source tool is designed for researchers and developers to explore and apply attention mechanisms within deep learning models, particularly in the context of multiple instance learning. It includes code for running MNIST-BAGS experiments and provides guidance for adapting the model to histopathology datasets like Breast Cancer and Colon Cancer. The implementation features a modified LeNet-5 model with Attention-based MIL pooling and uses the negative log-likelihood of the Bernoulli distribution as its objective function. It's a valuable resource for those looking to replicate or extend research in this specialized area of deep learning.