Research & Education

OpenNE is an open-source package designed for network embedding (NE) and serves as a comprehensive toolkit for network representation learning (NRL). It offers a standardized interface for both training and testing different NE models, ensuring scalability and flexibility. The package includes implementations of several typical NE models, such as DeepWalk, LINE, node2vec, GraRep, GCN, HOPE, GF, SDNE, and LE. A key feature is TADW, which allows for the incorporation of text attributes of nodes, enhancing the embedding process. OpenNE leverages TensorFlow, enabling GPU-accelerated training for improved performance. The toolkit also provides evaluation capabilities through node classification tasks, reporting Micro-F1, Macro-F1, and running time for various methods and datasets like Wiki and Cora.

RepDistiller is an open-source project that implements Contrastive Representation Distillation (CRD), as detailed in its ICLR 2020 paper. Beyond CRD, it serves as a comprehensive benchmark for 12 state-of-the-art knowledge distillation methods within the PyTorch framework. This tool enables users to run various distillation experiments, train student networks from pre-trained teacher models, and combine different distillation objectives. It provides scripts for fetching pre-trained teachers and running distillation processes, making it a valuable resource for researchers and developers working on model compression and efficiency.

Valence Labs, powered by Recursion, functions as an AI research engine dedicated to accelerating the mission of decoding biology to radically improve lives. The platform utilizes Recursion’s extensive data generation capabilities, including over 60 petabytes of phenomics and transcriptomics data, alongside its BioHive supercomputer for massive computing power. Valence Labs focuses on predicting cellular responses to perturbations, explaining molecular interactions, and discovering novel therapeutic hypotheses through lab-in-the-loop engines. This approach aims to build 'virtual cells'—mechanistic models of cellular function—to guide the discovery of new therapeutics and improve drug discovery outcomes.

ASE is a codebase designed for training reusable controllers for physically simulated characters, leveraging adversarial skill embeddings. It allows users to train models to imitate motion datasets, such as humanoids performing sword and shield actions, and to get up after falling. Beyond pre-training, ASE supports task-specific high-level controller training for various actions like reaching, heading, location, and striking. The codebase also includes an implementation of Adversarial Motion Priors (AMP) for imitating reference motions. It provides pre-trained models and tools for visualizing motion data, making it a valuable resource for researchers and developers in character animation and reinforcement learning.

意昂4 (EAON4) by Its IT Group specializes in smart fitness mirrors, integrating AI and motion capture technology to provide a personalized home fitness experience. The system offers real-time AI motion correction, identifying deviations in 0.3 seconds and providing audio-visual feedback to prevent injuries and enhance training effectiveness. Users benefit from over 1000 professional courses across 12 categories, designed by a team of top coaches. The AI generates personalized weekly training plans based on individual goals and fitness levels, adjusting difficulty as progress is made. The 55-inch 4K mirror seamlessly blends into home decor, and the system tracks multi-dimensional data, connecting with popular wearables to generate comprehensive fitness reports. It also fosters a community for user engagement and motivation.

Lila Sciences is building Scientific Superintelligence™ by combining advanced AI with autonomous labs to create the world's first Operating System for Science. This platform generates hypotheses, designs and runs experiments, and learns from new data in real time, significantly accelerating discovery across fields like advanced materials, therapeutics, chemicals, oil & gas, energy & environment, aerospace & defense, and biotech. The core technology, LILA's advanced AI model and proprietary AI Science Factory™ instruments, allows scientists to think, test, and learn at unprecedented speed and scale, outperforming traditional approaches in complex analysis and reasoning tasks. It aims to solve humankind's greatest challenges by executing the entire scientific method autonomously.

Peer AI is an advanced AI platform specifically designed for life sciences regulatory submissions. It significantly accelerates the regulatory path from molecule to medicine by enabling users to author submissions up to 95% faster. The platform offers a full lifecycle solution, providing one unified system from the first draft through submission, with end-to-end traceability for every AI decision. It features broadest document coverage with document-specific agents for pre-clinical, CMC, and clinical authoring. Peer AI also includes a Command Center for real-time program orchestration, identifying dependencies and bottlenecks, and predictive intelligence to anticipate reviewer queries before filing, ensuring stronger submissions from the outset. Built by regulatory experts, it integrates seamlessly with existing workflows and offers enterprise-grade security and compliance.

DeepHash is an open-source, lightweight deep learning library designed for hashing and quantization algorithms. It provides implementations of state-of-the-art deep hashing models such as DQN, DHN, DVSQ, DCH, and DTQ, with continuous updates and additions. The library is built to be extensible, actively encouraging researchers to contribute new deep hashing models based on its established framework. DeepHash is ideal for those working on efficient image retrieval and similarity search, offering tools and examples for data preparation, model training, and testing. It supports Python 3 and integrates with TensorFlow-GPU and OpenCV, making it suitable for technical users in academic or research settings.

CARLA is an open-source simulator specifically designed for autonomous driving research, offering a robust platform for the development, training, and validation of autonomous driving systems. It provides a rich set of open digital assets, including urban layouts, buildings, and vehicles, which can be freely used. The simulation platform is highly flexible, allowing for the specification of diverse sensor suites and environmental conditions to create realistic and varied scenarios. CARLA supports integration with various tools and frameworks, such as ROS-bridge for connecting to ROS, Scenario_Runner for executing traffic scenarios, and provides resources for reinforcement learning and conditional imitation learning models. It is built with Unreal Engine, currently supporting both UE 4.26 and UE 5.5 versions.

Gensaic is at the forefront of metabolic medicine, utilizing AI-enabled protein design to revolutionize therapeutic delivery. Their FORGE™ Platform employs Functional Optimization by Recursive Genetic Evolution, combining billion-scale protein evolution with AI-guided design to uncover novel mechanisms for tissue-selective intracellular delivery. This approach allows for the unbiased discovery of protein sequences that enable precise delivery to target tissues, specific cell types, and even subcellular compartments. Gensaic's focus is on developing potent, selective, and safe medicines for a broad range of metabolic indications, accelerating the discovery of next-generation therapies.

The Bamboo ViT-B16 Demo provides a practical demonstration of the Bamboo Vision Transformer (ViT) model's capabilities in the realm of computer vision. This tool allows users to interact with and understand how the ViT-B16 model processes and analyzes images. While the current live website indicates a build error, the underlying purpose is to showcase advanced image analysis techniques. It serves as a valuable resource for those interested in exploring the potential of transformer models in visual tasks, offering insights into their performance and applications.

ChemCrow is an AI tool built with Streamlit, designed to assist with chemistry research and AI-driven chemical tasks. It provides capabilities for chemical analysis and simulations, making it a valuable resource for professionals in the field. The tool aims to streamline complex chemical processes and data interpretation through artificial intelligence. While the specific features are not detailed, its focus on AI-assisted chemistry suggests advanced functionalities for scientific computing. The platform is currently paused on Hugging Face Spaces, indicating it may be in development or awaiting a restart by its creators.

Sciencessite is a versatile online platform designed to help users explore, learn, and grow their knowledge. It offers a range of tools and resources, including free educational quizzes to test knowledge, solved mathematics exercises for middle school students, and articles on various topics like the importance of science. The website emphasizes ease of use and clarity, aiming to provide valuable content without unnecessary complexity. It also features practical tools such as a global doctor/place search and information on financial services like Binance and Baridi Mob. The content is regularly updated to ensure accuracy and relevance, catering to a broad audience interested in learning and practical information.

Deep Spectral Segmentation is an AI tool designed for advanced image segmentation and spectral analysis. This tool is particularly beneficial for researchers and data scientists who work extensively with image data, providing capabilities to process and analyze visual information with deep learning techniques. It can be effectively utilized for developing sophisticated image processing applications, offering a robust platform for tasks that require detailed spectral insights. The tool is available as a Hugging Face Space, making it accessible for experimentation and integration into various projects.

Diception Demo is a generalist diffusion model designed for vision perception tasks. Hosted on Hugging Face Spaces, this tool allows users to upload an image and select from various tasks such as depth estimation, segmentation, or pose detection. For more advanced functionalities, users can optionally add specific points or categorize elements within the image. The tool then processes the input and displays detailed results as images. While the demo currently experiences a runtime error, its core functionality aims to provide a versatile platform for exploring and applying diffusion models in computer vision research and development.

GIRAFFE is an open-source project providing the code for the CVPR 2021 paper "GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields." This tool enables researchers and developers to explore 3D scene modeling and generative neural feature fields. It supports controllable image synthesis, allowing users to render images from trained models, including pre-trained options for datasets like Cars and CelebA-HQ. The repository also facilitates FID evaluation, training new networks from scratch, and implementing a 2D-GAN baseline. Users can adapt the tool for their own datasets by generating ground truth activations and adjusting image transformations, making it a valuable resource for advanced research in computer vision and machine learning.

Depth Anything Web is an AI-powered tool hosted on Hugging Face Spaces that provides real-time depth estimation from uploaded images. Users can easily submit an image file, and the application processes it to generate a detailed depth map, visually indicating which parts of the image are closer or farther away. This functionality is particularly useful for understanding spatial relationships within 2D images, offering a 3D-like perspective. The tool leverages the Xenova/depth-anything-small-hf model, making it a valuable resource for individuals involved in research, development, and educational pursuits within the fields of AI and computer vision. Its web-based interface ensures accessibility and ease of use for anyone looking to explore depth estimation without complex setups.

DiffVox is an AI-powered audio processing tool hosted on Hugging Face, designed to help users fine-tune vocal audio files. It provides a user-friendly interface with sliders to adjust various professional vocal effects, including equalization (EQ), compression, delay, and reverb. Users can customize their sound by tweaking principal components or by selecting from a range of pre-defined presets. This tool is ideal for those looking to experiment with and enhance vocal recordings, offering a flexible platform for audio exploration and modification. Its accessibility on Hugging Face makes it a convenient option for quick audio adjustments.

Dpt Depth Estimation + 3D is an AI tool designed to transform 2D images into interactive 3D models. Users can upload a standard 2D image, and the application processes it to generate a detailed depth map, which is then used to construct a 3D mesh. This 3D model can be viewed directly within the application and is available for download in the GLTF file format, making it compatible with various 3D software and platforms. The tool leverages DPT (Depth Prediction Transformer) technology to achieve accurate depth estimation, providing a straightforward solution for creating 3D assets from existing 2D visuals. It's particularly useful for those looking to quickly prototype 3D scenes or integrate 3D elements into their projects without extensive 3D modeling experience.

Emotions is a unique AI tool that enables users to interact with and control the emotional expressions of a Reachy Mini robot. Through its intuitive Emotions Wheel app, users can browse and select from more than 138 pre-defined robot behaviors, each organized by emotion colors. Clicking a badge instantly makes the robot move, allowing for real-time adjustment of its emotional state. This platform is ideal for exploring human-robot interaction, understanding emotional responses in robotics, and developing engaging robot behaviors. It provides a hands-on experience for both enthusiasts and developers interested in the expressive capabilities of robots.

Esmfold is an AI-powered tool designed for predicting protein structures, offering a valuable resource for the scientific community. It enables biologists, biochemists, and researchers to accurately model and analyze complex protein structures, which is crucial for understanding biological functions and developing new therapies. The tool is available for free, making advanced molecular modeling accessible for scientific research and educational purposes. While the specific features are not detailed, its core utility lies in providing structural insights into proteins, aiding in drug discovery, and enhancing academic studies in molecular biology.

Embedding Converter is a specialized AI utility tool designed to facilitate the conversion of SD1.5 embedding files into a compatible SDXL format. This tool is particularly useful for developers and data scientists working with AI models, enabling them to seamlessly integrate older embedding files into newer SDXL-based projects. Hosted on Hugging Face Spaces and built with Gradio, it offers a straightforward process: users upload their SD1.5 embedding file, the application processes it, and then provides the converted version for download. This ensures compatibility and efficiency in AI development workflows, making it easier to leverage existing resources within updated frameworks.

FateZero Inference is an AI tool designed for model inference, providing capabilities for both model deployment and research. Hosted on Hugging Face Spaces, it aims to facilitate the use and exploration of AI models within the community. While the platform experienced a runtime error at the time of scraping, its intended purpose is to offer a space for users to interact with and utilize AI models. It is positioned as a free resource, making it accessible for individuals and teams involved in AI development and academic research.

Fathom DeepResearch is an AI-powered tool designed to streamline the research process. Users can enter a text question, and the application automatically conducts web searches and utilizes various other tools to gather relevant information. The core functionality lies in its ability to synthesize this retrieved data and present a clear, concise answer, accompanied by a preview of the source material. This makes it particularly useful for quickly obtaining summarized information from diverse online sources, leveraging fathom search and synthesizer models for deep research capabilities. It is hosted on Hugging Face Spaces by Fractal AI Research.