Research & Education

DeepEMD offers a PyTorch implementation for few-shot image classification, based on the research paper "DeepEMD: Few-Shot Image Classification with Differentiable Earth Mover's Distance and Structured Classifiers." This tool is designed to address the challenge of learning from limited labeled data by employing the Earth Mover's Distance (EMD) as a metric for structural matching between image regions. It includes a cross-reference mechanism to mitigate issues from cluttered backgrounds and intra-class variations, and supports k-shot classification through a structured fully connected layer. DeepEMD has demonstrated significant performance improvements on benchmarks like miniImageNet, tieredImageNet, FC100, and CUB, without requiring extra training or testing data. The repository provides code for model pre-training, meta-training, and evaluation, along with options for different EMD solvers and model configurations.

EducUp Studio is a platform designed for educators to create and monetize their knowledge through interactive, gamified asynchronous courses. It enables the transformation of traditional learning materials into engaging educational content, aiming to boost student engagement and knowledge retention. The platform supports educators in establishing a strong online presence and expanding their educational reach by offering tools for course creation and monetization. It focuses on making education accessible and interactive, covering subjects like English, Math, Digital Marketing, and Personal Finance, and is suitable for various educational contexts including SAT, ACT, and GED preparation.

DeepRL-Tutorials is an open-source repository offering high-quality implementations of various Deep Reinforcement Learning (DRL) algorithms, primarily written in PyTorch. The project emphasizes readability and understanding, making it an excellent resource for those looking to learn and practice DRL concepts. It includes implementations of algorithms such as DQN, Double DQN, Dueling DQN, Rainbow, A2C, PPO, and more, each accompanied by relevant research papers. The tutorials are presented as IPython Notebooks, providing a structured way to explore and experiment with these advanced AI techniques. It requires Python 3.6, Numpy, Gym, Pytorch 0.4.0, Matplotlib, and OpenCV.

F0lkl0r3.dev is a unique digital archive that brings the rich history of computing to life through oral history interviews from the Computer History Museum. This platform enriches these invaluable firsthand accounts with AI-generated context, relevant visuals, and interconnected links, creating a searchable and interlinked map of computing history. It serves as an essential resource for historians, researchers, students, and anyone with a keen interest in the evolution of technology. By making complex historical narratives more accessible and engaging, F0lkl0r3.dev allows users to explore the stories of the pioneers who shaped the digital world, understand the intricate connections between various innovations, and gain deeper insights into the foundational moments of computer science.

Diffusion-Explorer is an interactive tool designed to communicate the geometric intuitions behind diffusion and flow-based generative models. It offers key functionality such as implementing various training objectives like Flow Matching and Denoising Score Matching. Users can observe the dynamics of generated samples over time for pretrained models, see how samples evolve through training, and even train models on custom hand-drawn distributions. The project also includes a Rectified Flow Explainer, an interactive blog post with animated visualizations demonstrating how flow matching learns curved trajectories, why curved paths are problematic for few-step sampling, and how rectified flow iteratively straightens trajectories. This tool is currently a work in progress and is mainly educational.

neural-combinatorial-rl-pytorch offers a PyTorch implementation of Neural Combinatorial Optimization with Reinforcement Learning, based on the research paper. This open-source tool provides a basic RL pretraining model that utilizes greedy decoding. A notable feature is its use of an exponential moving average critic instead of a traditional critic network, which has been shown to significantly improve results, particularly for the Traveling Salesperson Problem (TSP). The implementation supports a stochastic decoding policy during training and beam search for testing. It currently includes support for a sorting task and the planar symmetric Euclidean TSP, with clear guidelines for extending it to other combinatorial optimization problems by providing a dataset class and a reward function. The repository also details dependencies and provides performance results for both TSP and sorting tasks, demonstrating its generalization capabilities.

The Embedded-Engineering-Roadmap is an open-source resource designed to assist both aspiring and current Embedded Systems Engineers in navigating their career path and expanding their skill sets. It offers a comprehensive guide structured into three fundamental areas: Software, Hardware, and Soft Skills. The roadmap provides a curated list of learning resources, categorized by type (books, videos, articles, links) and quality (beginner-friendly, invaluable/comprehensive). It emphasizes hands-on projects as the most effective learning approach and includes links to various project ideas and educational websites. Additionally, it offers guidance on career development, search strategies, and recommended IDEs and VS Code extensions, making it a valuable tool for anyone looking to excel in embedded systems.

Emotion-LLaMA is an advanced open-source AI model designed for multimodal emotion recognition and reasoning, leveraging instruction tuning. It addresses the limitations of traditional single-modality approaches by seamlessly integrating audio, visual, and textual inputs through emotion-specific encoders. The model aligns features into a shared space and employs a modified LLaMA model, significantly enhancing both emotional recognition and reasoning capabilities. It was accepted at NIPS 2024 and has achieved top scores in various challenges, including the MER2024 Challenge. The project also includes the MERR dataset, which contains a large number of coarse-grained and fine-grained annotated samples across diverse emotional categories, enabling models to learn from varied scenarios and generalize to real-world applications.

Eagle 2.5 is a family of frontier vision-language models (VLMs) developed by NVlabs, specifically engineered for long-context multimodal learning. Unlike many existing VLMs that focus on short-context tasks, Eagle 2.5 excels at challenges like long video comprehension and high-resolution image understanding, providing a generalist framework for both. It supports up to 512 video frames and is trained jointly on image and video data, including the novel Eagle-Video-110K dataset. Key innovations include Information-First Sampling for optimal image and text retention, Progressive Mixed Post-Training for enhanced context length processing, and Diversity-Driven Data Recipe. The model also features significant efficiency and framework optimizations, such as GPU memory optimization and inference acceleration, making it suitable for advanced research and development in multimodal AI.

Lenzy AI offers a comprehensive analysis and visualization of two decades of Hacker News discussions. Utilizing clustering algorithms, the platform identifies and presents key trends, recurring patterns, and community insights from the vast dataset. This tool is designed for researchers and analysts to explore the evolution of technology conversations, pinpoint dominant themes, and understand the collective interests of the developer community over a significant period. It provides an overview of discussed topics, making it valuable for anyone interested in the historical trajectory of tech discourse on Hacker News.

Face_Pytorch offers an open-source implementation of various face recognition algorithms within the PyTorch framework. This project includes well-known algorithms such as ArcFace, CosFace, and SphereFace, providing a comprehensive toolkit for researchers and developers. It supports data preparation for CNN training using datasets like CASIA-WebFace and Cleaned MS-Celeb-1M, aligned by MTCNN. The project also facilitates performance testing on benchmarks like LFW, AgeDB-30, CFP-FP, and MegaFace, with detailed verification results provided for different model types and protocols. It's designed for those looking to implement and evaluate face recognition models, offering flexibility for custom dataset paths and parameters.

FishNet offers the implementation code for the FishNet architecture, a versatile backbone designed for image, region, and pixel-level prediction tasks. Based on a NeurIPS 2018 paper, this tool provides pre-trained models with varying parameters and FLOPs, including FishNet99, FishNet150, and FishNet201, with reported Top-1 and Top-5 accuracies. It supports training with PyTorch and includes configurations for data augmentation methods like random flip, random crop, and random PCA lighting. The project also details how to load and utilize these models, making it a valuable resource for researchers and developers working on computer vision challenges.

FaceRecognitionApp is an open-source Android application developed by Kristian Lauszus in 2016, designed to showcase face recognition capabilities. The app implements Eigenfaces and Fisherfaces algorithms for facial recognition, leveraging the FaceRecognitionLib library for its core calculations. It provides a practical example for developers interested in integrating face recognition into Android applications. The project is released under the GNU General Public License, encouraging community contributions and modifications. It requires Android Studio, the Android NDK, OpenCV Android SDK, and Eigen3 libraries for building and running, with detailed instructions provided for both basic and advanced users who wish to modify the source code.

ExtremeNet is an open-source object detection system that employs a bottom-up approach to identify objects within images. It achieves this by detecting four extreme points (top-most, left-most, bottom-most, right-most) and one center point of objects using a standard keypoint estimation network. These five keypoints are then grouped into a bounding box if they are geometrically aligned. This method transforms object detection into a purely appearance-based keypoint estimation problem, bypassing region classification or implicit feature learning. The project is built upon the CornerNet code and integrates code from Deep Extreme Cut (DEXTR) for instance segmentation, allowing it to generate coarse octagonal masks and further refine them for improved Mask AP. It provides code for training, evaluation, and demo purposes, supporting benchmark evaluation on datasets like MS COCO.

FSDrive is the official implementation for the research paper "FutureSightDrive: Thinking Visually with Spatio-Temporal CoT for Autonomous Driving," which was recognized as a NeurIPS 2025 spotlight. This tool introduces a novel spatio-temporal Chain-of-Thought (CoT) approach, allowing end-to-end autonomous driving Visual Language Agents (VLA) to visually process and plan trajectories. It uniquely unifies visual generation and understanding with minimal data, marking a significant advancement in applying visual reasoning to autonomous driving. FSDrive provides comprehensive instructions for installation, data preparation, training, inference, evaluation, and visualization, making it a valuable resource for researchers and developers in the autonomous driving domain.

The Knight in the Panther's Skin is a digital edition of Shota Rustaveli's 12th-century Georgian epic poem, Vepkhistkaosani. This interactive platform provides the full text in English (Wardrop 1912 translation) with a parallel Georgian text from the critical edition, allowing for a bilingual reading experience. Users can explore all 47 chapters, access detailed annotations, and view illustrations that bring the allegorical masterpiece to life. The tool enhances engagement with this historically significant work by weaving themes of friendship, love, and devotion into an accessible digital format, making it ideal for students, scholars, and enthusiasts of medieval literature and Georgian culture.

Lyspeak, despite its name suggesting a language learning tool, functions as an affiliate website in Turkish, focusing on online betting and casino bonuses. The site provides lists of current promotions, such as welcome bonuses, free spins, and no-deposit bonuses, from numerous gambling platforms like Roketbet, Betmoney, Milyar, and Fikstürbet. It aims to guide users to reliable and licensed sites offering these bonuses, detailing different types of bonuses like investment bonuses, loss bonuses, and freebet offers. The platform also includes an FAQ section addressing common questions about bonuses, their usage, and terms and conditions, positioning itself as a resource for individuals interested in online gambling promotions.

Qwen-VL, developed by Alibaba Cloud, is a powerful open-source large vision language model (LVLM) that accepts image, text, and bounding box inputs, and outputs text and bounding boxes. It offers strong performance, significantly surpassing existing open-sourced LVLMs on multiple English evaluation benchmarks. Key features include multi-lingual support for English, Chinese, and multi-lingual conversations, end-to-end recognition of bi-lingual text in images, and multi-image interleaved conversations. It is also the first generalist model to support grounding in Chinese, allowing for bounding box detection through open-domain language expression. The model boasts fine-grained recognition and understanding with a 448x448 resolution, promoting detailed text recognition and document QA.

prettygraph is a Python-based web application developed by @yoheinakajima, designed to demonstrate a new UI pattern for text-to-knowledge graph generation. While it's an experimental project and not intended as a robust framework, it provides a simple yet interactive way to visualize knowledge graphs. The application uses Flask for the backend, LiteLLM for generating predictions that transform text inputs into JSON formatted graph data, and Cytoscape.js for visualization. A key feature is its dynamic UI, where the graph regenerates and updates in real-time with each period insertion in the text input, offering color-coded nodes and edges for better visual distinction. It requires an OpenAI API key for operation.

GeoChat is an open-source, grounded Large Vision Language Model (LVLM) specifically designed for Remote Sensing (RS) applications. Unlike general-domain models, GeoChat is tailored to handle high-resolution RS imagery and employs region-level reasoning for detailed scene interpretation. It leverages a newly created RS multimodal dataset and is fine-tuned using the LLaVA-1.5 architecture, resulting in robust zero-shot performance across various RS tasks. These tasks include image and region captioning, visual question answering, scene classification, visually grounded conversations, and referring object detection. GeoChat also introduces a novel data generation pipeline to create rich instruction sets for the RS domain, making it a valuable tool for researchers and developers in AI and remote sensing.

ONCETALK is an advanced AI tool engineered for dynamic and intelligent conversations. It leverages real-time internet data to ensure responses are always up-to-date and accurate, making it a reliable source for current information. The platform continuously learns and adapts, improving its conversational capabilities over time. This adaptability makes ONCETALK suitable for a wide array of information retrieval and interactive dialogue tasks across various domains. By offering contextually relevant and evolving insights, ONCETALK significantly enhances user engagement, providing a more intelligent and responsive interaction experience. Its core strength lies in its ability to process and utilize live data, setting it apart in delivering timely and precise information.

RoboVerse is an open-source initiative providing a unified platform, dataset, and benchmark specifically designed for scalable and generalizable robot learning. It aims to accelerate research and development in robotics and AI by offering a comprehensive ecosystem for creating, testing, and evaluating robot learning algorithms. The platform integrates various simulation frameworks and renderers, including Isaac Lab, Isaac Gym, MuJoCo, and Blender, alongside data from projects like RLBench and Maniskill. RoboVerse encourages community contributions and provides detailed documentation and tutorials to help users get started. Its focus on a standardized environment and extensive datasets makes it a valuable resource for advancing the field of robot learning.

Hyperspectral-Image-Super-Resolution-Benchmark is an open-source collection of resources dedicated to hyperspectral image super-resolution. Curated by Junjun Jiang, this benchmark provides a comprehensive list of techniques and papers for generating high spatial and high spectral resolution images. It covers four main classes of super-resolution: spatiospectral super-resolution (SSSR), spectral super-resolution (SSR), single hyperspectral image super-resolution (SHSR), and multispectral image and hyperspectral image fusion (MHF). The resource includes pioneer work, technique reviews, and recent advancements, often with links to PDF papers and code, making it an invaluable tool for researchers and academics in the field.

Research in English News is a dedicated platform designed to make the latest academic research accessible to a broader audience. It translates complex scientific papers into concise, easy-to-understand articles, covering a wide range of topics from astrophysics and quantum communication to mental health and computer vision. The website features summaries of groundbreaking studies, highlighting key findings and their implications, such as new AI models for retinal scans, advancements in autonomous system safety, and insights into strange metals. This approach democratizes access to cutting-edge scientific knowledge, allowing individuals to stay informed about significant developments without needing to navigate dense scholarly content.