Research & Education

Street-fighter-ai is an AI agent specifically designed and trained using deep reinforcement learning to play the classic game "Street Fighter II: Special Champion Edition." The agent operates by making decisions based solely on the RGB pixel values of the game screen, demonstrating a sophisticated approach to game AI. It has been shown to achieve a 100% win rate in the first round of the final level, though this can involve overfitting. The project provides detailed instructions for environment setup, running tests with pre-trained models, and even training your own models. It leverages open-source libraries like OpenAI Gym Retro and Stable-Baselines3, making it a valuable resource for researchers and enthusiasts in AI and reinforcement learning.

T-Rex2 is an advanced object detection model developed by IDEA-Research, designed to overcome the limitations of traditional, closed-set object detection systems. By integrating both text and visual prompts, T-Rex2 harnesses the strengths of both modalities, providing robust zero-shot capabilities. This makes it a versatile tool for identifying and locating objects within images across a wide range of applications, including agriculture, industry, livestock monitoring, biology, medicine, OCR, retail, electronics, transportation, and logistics. It supports three main workflows: interactive visual prompt, generic visual prompt, and text prompt, covering most object detection scenarios. The project provides API access and a local Gradio demo for easy implementation and experimentation.

TextRank is a Python implementation of the TextRank algorithm, specifically designed for automatic keyword and sentence extraction, which facilitates summarization. This particular implementation distinguishes itself by utilizing Levenshtein distance to determine the relationship between text units, offering a unique approach to text analysis. The project is based on the foundational paper "TextRank: Bringing Order into Text" by Rada Mihalcea and Paul Tarau. It provides functionalities for both keyword and sentence extraction, making it a valuable tool for researchers and developers working with text data. The library is installable via pip and requires NLTK resources, which can be fetched using a simple command.

tf-cpn is a Tensorflow re-implementation of the Cascaded Pyramid Network (CPN), a state-of-the-art model for multi-person pose estimation that won the 2017 COCO Keypoints Challenge. This open-source tool provides researchers and developers with the code and pre-trained models necessary to implement and experiment with advanced pose estimation. It includes detailed instructions for training on the MSCOCO dataset, downloading base models, and running validation tests. The repository also offers pre-trained models for various configurations (ResNet-50, ResNet-101 with different input sizes) and provides performance metrics on COCO minival and test-dev datasets, making it a valuable resource for academic and practical applications in computer vision.

text_gcn is an open-source implementation of Graph Convolutional Networks (GCNs) specifically designed for text classification tasks. This tool provides the necessary code to reproduce the results presented in the paper "Graph Convolutional Networks for Text Classification" from the AAAI 2019 conference. It requires Python 2.7 or 3.6 and Tensorflow >= 1.4.0, making it accessible for those familiar with these environments. The repository includes scripts for data preparation, graph building, and model training, along with examples for various datasets like 20ng, R8, R52, ohsumed, and mr. An inductive version, fast_text_gcn, is also available for scenarios where test documents are not included in the training process.

ThoughtSource is an open and central resource designed for researchers and developers working with chain-of-thought reasoning in large language models. It provides a comprehensive collection of datasets, including general question answering, scientific/medical QA, and math word problems, all formatted for standardized chain-of-thought analysis. The platform also includes tools for generating reasoning chains with various language models (OpenAI, Hugging Face) and evaluating their performance. With its dataset annotator and viewer applications, ThoughtSource aims to foster a community around improving trustworthy and robust reasoning in AI, particularly for scientific research and medical practice. It is developed by the Samwald research group.

Findsight AI is a sophisticated search engine designed for exploring and comparing core ideas across a vast collection of non-fiction works. It facilitates syntopical reading by allowing users to discover and compare claims made by various sources, observe how authors approach different issues, and navigate related claims to build their own learning paths. The platform offers both basic and AI-powered filters to refine search results. Basic filters include 'mention' for literal text searches and 'references' for named entities like skills or concepts. AI-powered filters, such as 'state' and 'answer', enable advanced searching by allowing users to find related claims based on a custom input or identify claims that address a specific question. Users can also find links to original books or articles, making it a valuable resource for academic research and in-depth study.

LOTUS Normal is an AI tool designed to generate high-quality predictions from input images, offering both generative and discriminative outputs. This application allows users to upload an image and optionally specify a seed number for generation. While the tool's specific functionalities beyond image prediction are not detailed, its presence on Hugging Face suggests it leverages advanced machine learning models for its operations. The platform itself, Hugging Face, provides various pricing tiers for its services, including storage and compute resources for running such applications, indicating that while the core tool might be accessible, underlying infrastructure costs could apply for extensive use.

LookinGlassRGBD is an AI tool designed for processing RGBD (Red, Green, Blue, Depth) images, facilitating advanced 3D scene understanding. It allows users to analyze depth information alongside color data, which is crucial for applications requiring precise spatial awareness. The tool is particularly beneficial for researchers and developers in the computer vision field, offering capabilities for tasks such as object recognition, environmental mapping, and robotic navigation. Hosted on Hugging Face Spaces, it leverages community-driven machine learning models, providing a platform for experimentation and development in 3D computer vision.

LoRA Roulette is an innovative AI tool hosted on Hugging Face that allows users to explore the creative potential of combining different LoRA (Low-Rank Adaptation) models. The application generates unique images by randomly selecting and blending two LoRA models, which users can then influence with a custom text prompt. It provides functionalities to shuffle the selected models and adjust their individual weights or influence on the final output, offering a dynamic way to experiment with various AI art styles and characteristics. This tool is ideal for artists, researchers, and enthusiasts looking to understand the interplay of different LoRA models and generate novel visual content.

MathGPTPro, rebranded as Mathos AI, is a comprehensive AI-powered platform designed to assist students with a wide range of mathematical and scientific problems. It functions as an AI math solver, calculator, and tutor, capable of handling subjects from algebra and calculus to physics and engineering. Users can input problems by snapping a photo, speaking, or typing, including LaTeX. The tool provides detailed, step-by-step solutions and offers features like an AI Tutor that coaches in real-time, a PDF Homework Helper to solve entire worksheets, interactive games for concept reinforcement, and study tools to generate quizzes, flashcards, and video explainers from solutions. Mathos AI aims to make math learning more engaging and accessible for over 5 million students globally.

MPLUG Owl2 is an AI tool hosted on Hugging Face, providing a platform to explore and test the mPLUG-Owl2 model. This tool is designed for users interested in experimenting with advanced AI models, particularly within the domain of open-source development and research. While the live website currently displays a runtime error, indicating a temporary issue with the application, its intended purpose is to offer access to the mPLUG-Owl2 model for various applications. It is available for free, making it accessible for educational and research purposes, allowing individuals to delve into the capabilities and potential of this specific AI model.

MNIST Adversarial is an AI tool hosted on Hugging Face Spaces, designed for exploring and understanding adversarial attacks on machine learning models, specifically those trained on the MNIST handwritten digit dataset. Users can generate adversarial examples that are crafted to fool the classification model, providing insights into the vulnerabilities and robustness of AI systems. This tool is particularly valuable for educational purposes, allowing students and researchers to experiment with and visualize the effects of small, imperceptible perturbations on model predictions. While the tool itself is free to access, the underlying Hugging Face platform offers various paid plans for enhanced compute resources and features, making it suitable for both casual exploration and more intensive research.

MobileSAM is an AI tool designed for efficient image segmentation, specifically optimized for deployment on mobile devices. This allows for on-device image processing tasks, reducing the need for constant cloud connectivity and enhancing privacy. Developers can leverage MobileSAM to build AI-powered mobile applications that require robust image analysis capabilities directly on the user's device. The tool is hosted on Hugging Face Spaces, providing a platform for community collaboration and deployment. While the core functionality is image segmentation, the underlying Hugging Face platform offers various pricing tiers for compute resources, catering to different scales of application development and usage.

MultiMAE is an AI tool available on Hugging Face Spaces that demonstrates image reconstruction using a masking approach. Users can upload an image and interactively control the percentage of visible parts, allowing them to observe how the MultiMAE model reconstructs the masked areas. This provides a clear visualization of the model's understanding and generative capabilities in computer vision. It is particularly useful for researchers and developers interested in understanding and experimenting with image reconstruction techniques and masked autoencoders. The tool offers a hands-on experience to explore the impact of varying mask percentages on the quality and coherence of the reconstructed image.

Music Descriptor is an AI-powered application hosted on Hugging Face that offers comprehensive music analysis. Users can upload audio files or record live music to receive detailed insights into its characteristics. The tool identifies various aspects of music, including genres, instruments present, and the emotional content conveyed. It then provides a breakdown of top predictions for each category, making it a valuable resource for understanding musical compositions. This tool is designed for anyone interested in a deeper analysis of music, from casual listeners to professionals.

Music Flamingo is an AI-powered tool hosted on Hugging Face that enables users to deeply analyze music. By simply uploading an audio file or providing a YouTube video link, users can then pose various questions about the music. The tool is designed to extract audio and provide detailed insights into aspects such as genre, tempo, lyrics, chords, or even a comprehensive analysis of the musical composition. This makes it a versatile platform for anyone looking to understand the intricacies of a piece of music without requiring specialized musical knowledge.

Open RL Leaderboard is a platform designed for comparing and evaluating reinforcement learning (RL) algorithms. It serves as a centralized hub where researchers and practitioners can benchmark their RL agents against others in the community. The leaderboard facilitates the tracking of progress in RL research and helps identify top-performing methods. By providing a standardized evaluation framework, it supports the advancement of the RL field. The platform appears to be hosted on Hugging Face Spaces, indicating a focus on community collaboration and accessibility for machine learning projects. While currently experiencing a runtime error, its core purpose is to foster transparency and competition in RL development.

The Nucleotide Transformer Benchmark is a specialized tool designed for evaluating the performance of DNA foundational models across various downstream tasks. Hosted on Hugging Face Spaces by InstaDeepAI, this application allows researchers to generate leaderboards by selecting specific tasks and metrics. It provides a clear overview of how different models perform, making it an invaluable resource for benchmarking and analysis in the fields of bioinformatics and genomics research. The tool facilitates direct comparison of transformer models, aiding in the advancement and understanding of AI applications in nucleotide sequence data.

paper-central is an AI-powered tool hosted on Hugging Face designed to streamline the process of browsing and discovering academic papers. Users can efficiently navigate through a vast collection of research by applying various filters, including publication dates, specific categories, conferences, authors, or keywords found in paper titles. The tool presents matching results in an interactive table, making it easy to review and select papers of interest. This functionality is particularly useful for researchers, students, and academics who need to conduct literature reviews, stay updated on new publications, or find relevant studies for their work, offering a more organized approach to academic paper exploration.

PROTEIN GENERATOR is an AI-powered tool hosted on Hugging Face that facilitates the design of novel proteins. Users can define protein characteristics by specifying a desired length, inputting a custom amino acid sequence, or selecting a structural motif. The tool offers advanced customization options, allowing users to introduce biases for secondary structure elements, control amino-acid composition, or adjust hydrophobicity. This functionality makes it a valuable resource for researchers and academics involved in protein engineering, drug discovery, or synthetic biology, providing a flexible platform for exploring protein design principles and generating new protein candidates for various applications.

ProteinMPNN + ESM is an AI-powered tool hosted on Hugging Face that facilitates the prediction of novel protein sequences designed to fold into a specified protein structure. This application allows users to either provide a PDB code for an existing protein structure or upload their own protein structure file. It integrates the capabilities of ProteinMPNN for sequence design with ESM for enhanced protein understanding. Users can customize settings such as the number of sequences to generate, making it a versatile tool for researchers in structural biology and protein engineering who need to explore sequence-structure relationships.

The Thousand Brains Project is an open-source initiative dedicated to rethinking AI from the ground up, drawing inspiration from the neocortex. It aims to build a general-purpose system for modeling sensorimotor data, rather than static datasets, making it highly applicable to robotics and even abstract sensorimotor tasks like navigating the web or processing language. The project emphasizes energy and data efficiency, continuous learning, and modularity, allowing for flexible architectures. Unlike current AI approaches, it focuses on learning through active interaction with the world, similar to how humans learn. The project is funded by Jeff Hawkins and the Gates Foundation, and it provides open-source code, research meeting recordings, and documentation for community involvement.

Deep Research with Google Gemini is an AI-powered application hosted on Hugging Face Spaces, designed to facilitate in-depth exploration of any subject. Users can input their research topics and optionally include local resources, after which the system leverages Google Gemini to analyze and gather information. This tool is ideal for anyone needing to conduct thorough research, providing a structured approach to information retrieval and analysis. It aims to streamline the research process by offering a platform where questions can be posed and comprehensive data collected efficiently, making it a valuable asset for academic and professional researchers alike.