Research & Education

stat212b is a comprehensive open-source repository on GitHub, offering course materials for a Deep Learning Topics Course from UC Berkeley, taught by Joan Bruna. The curriculum is divided into three main parts: Convolutional Neural Networks, Deep Unsupervised Learning, and Miscellaneous Topics. It covers advanced concepts such as invariance, stability, variability models, scattering extensions, and various types of autoencoders and generative adversarial networks. The repository includes lecture PDFs, reading lists, and guest lectures from prominent researchers like Wojciech Zaremba and Soumith Chintala. This resource is ideal for students and researchers looking to delve into the theoretical and practical aspects of deep learning.

STAT 453: Intro to Deep Learning @ UW-Madison (Spring 2020) is an open-source GitHub repository offering comprehensive course materials for an introductory deep learning class. The repository includes lecture notes, assignments, and code examples covering fundamental concepts such as single-layer neural networks, linear algebra for deep learning, gradient descent, and PyTorch. It also delves into advanced topics like multilayer perceptrons, regularization, convolutional neural networks (CNNs), recurrent neural networks (RNNs), autoencoders, and generative adversarial networks (GANs). This resource is ideal for students and educators looking for structured content to learn or teach deep learning and generative models.

d2l-en is an interactive deep learning book designed to make deep learning approachable through hands-on learning. The entire book is drafted in Jupyter notebooks, seamlessly integrating exposition, figures, math, and interactive examples with self-contained code. It offers sufficient technical depth to serve as a starting point for aspiring applied machine learning scientists and includes runnable code to demonstrate practical problem-solving. The resource is open-source, allowing for rapid updates by both the authors and the community, and is complemented by a forum for technical discussions and questions. Adopted by over 500 universities in 70 countries, including Stanford, MIT, Harvard, and Cambridge, d2l-en is a highly regarded educational tool.

deep-image-prior is an open-source project that offers a novel approach to image restoration using neural networks, notably without requiring a traditional learning phase. It leverages the inherent structure of convolutional neural networks as a prior for image reconstruction. The repository provides Jupyter Notebooks that allow users to reproduce figures and experiments from the associated 'Deep Image Prior' CVPR 2018 paper. This includes notebooks for tasks like denoising, inpainting, super-resolution, and activation maximization. Users should be aware that optimization may not converge on some GPUs, and it's recommended to verify results against the paper's findings, potentially by adjusting precision settings or disabling cudnn.

SymbolicAI is a neuro-symbolic framework designed to integrate classical Python programming with the programmable nature of Large Language Models (LLMs). It emphasizes a modular and extensible design, allowing users to easily create custom engines, host local models, and interface with external tools like web search or image generation. The framework introduces 'Symbol' objects, which can operate in either syntactic (normal Python value) or semantic (neuro-symbolic engine-wired) modes, enabling complex chains of operations. A key differentiator is its implementation of Design by Contract principles for LLMs, helping to build correctness directly into the design through decorators, data models, and validation constraints to mitigate hallucination.

Sudoku is an open-source project hosted on GitHub that explores the application of convolutional neural networks (CNNs) to solve Sudoku puzzles. The project showcases a computational method for tackling this popular number puzzle, which involves filling a 9x9 grid with digits such that each row, column, and 3x3 subgrid contains all digits from 1 to 9. It provides a dataset of 1 million generated Sudoku games for training and includes Python scripts for generating puzzles, training the model, and testing its performance. The model, consisting of 10 blocks of convolution layers, achieves an accuracy of 0.86 in solving Sudoku puzzles, demonstrating the potential of simple CNNs without rule-based postprocessing. This project is valuable for researchers and students interested in AI, machine learning, and problem-solving.

TAADpapers is a meticulously curated list of essential academic papers focusing on textual adversarial attacks and defenses in natural language processing (NLP). Maintained by researchers at UChicago and THUNLP, this resource is invaluable for academics, researchers, and professionals working in NLP, machine learning, and cybersecurity. The collection is organized into key sections including toolkits, survey papers, various levels of attack papers (sentence, word, char, multi-level), defense papers, certified robustness, and benchmarks. Each entry typically includes links to the paper's PDF, and often to associated codebases or websites, making it a practical and accessible reference for staying updated on the latest advancements and challenges in robust NLP.

TextMatch is a comprehensive open-source library designed for various natural language processing tasks, including semantic matching, text classification, text embedding, text clustering, and text retrieval. It provides an easy-to-use framework for training models and exporting representation vectors. The library supports a wide array of models and techniques, ranging from traditional methods like Bow, TFIDF, and Ngram-TFIDF to advanced deep learning models such as BERT, ALBERT, and SimCSE. Additionally, it incorporates algorithms for clustering (Kmeans, DBSCAN), dimensionality reduction (PCA), and efficient similarity search (FAISS). TextMatch is ideal for developers and researchers looking to implement and experiment with different text processing and matching algorithms.

PrecedentAI is an advanced AI-powered legal research tool designed to streamline the legal research process for professionals. It allows users to instantly navigate through a vast database comprising millions of legal cases, briefs, and articles. The platform is engineered to provide precise answers and relevant citations, consolidating all necessary information into a single, efficient search. By leveraging artificial intelligence, PrecedentAI aims to significantly reduce the time and effort typically associated with complex legal inquiries, helping users quickly find the information they need to support their legal arguments and decisions.

Discuria is an AI-powered platform designed for researchers and academics to discover, engage with, and discuss scientific papers. It provides access to over 200 million papers from major databases like arXiv, Semantic Scholar, OpenAlex, CrossRef, and PubMed. Users can search for papers directly or ask questions, with the AI understanding their needs. The platform allows for inline annotations, AI-powered chat for summaries and explanations, and threaded discussions with other researchers. Additionally, users can upload any PDF to annotate and share, and utilize a read-aloud feature with synchronized word highlighting, making complex research more accessible and interactive.

Text-Summarization-Papers offers an exhaustive, open-source list of academic papers focused on text summarization. Curated from eight top conferences including ACL, EMNLP, and NeurIPS, the collection spans from 2013 to 2020. Users can leverage a paper retrieval system to easily find highly-cited papers, track the latest research, and identify milestone works for beginners. The resource also categorizes papers by research concepts, highlighting hot topics and trends since 2019, such as scientific paper-based summarization, new datasets, pretrained models, graph neural networks, and factuality evaluation. It encourages community contributions for updating papers, concepts, and datasets.

tf-dann is an open-source implementation of Domain-Adversarial Neural Networks (DANN) in Tensorflow, designed to address domain adaptation challenges. It leverages a gradient reversal layer to enable unsupervised domain adaptation through backpropagation, allowing models to generalize effectively across different domains even without labeled data in the target domain. The repository includes practical examples, such as experiments on a simple Blobs dataset and a recreation of the MNIST experiment from the original DANN papers. It provides instructions for generating the synthetic MNIST-M dataset and details the implementation of the `flip_gradient` operation using `tf.gradient_override_map`. This tool is ideal for researchers and developers working on machine learning models that need to perform well across varied data distributions.

tf-rnn-attention provides a Tensorflow implementation of the attention mechanism specifically designed for text classification tasks. This open-source project is inspired by the research presented in "Hierarchical Attention Networks for Document Classification" by Zichao Yang et al. It serves as a valuable resource for developers and researchers looking to integrate attention mechanisms into their natural language processing models. The repository includes Python code for attention, training, and utility functions, along with a visualization example. Users can leverage this tool to build and experiment with text classification models that benefit from the interpretability and performance enhancements offered by attention mechanisms.

TTS-arxiv-daily is an open-source project designed to automatically track and update the latest Text-to-Speech (TTS) papers from arXiv. Utilizing GitHub Actions, the repository updates every 12 hours, ensuring that users have access to the most current research in the field. This tool is particularly valuable for researchers, developers, and academics who need to stay abreast of new developments in speech synthesis without manually sifting through arXiv. It streamlines the process of literature review, making it easier to identify and access relevant publications quickly. The project is hosted on GitHub, providing transparency and community collaboration opportunities.

TxAgent is an AI agent developed for precision therapeutics, focusing on generating personalized treatment recommendations. It employs multi-step reasoning and real-time biomedical knowledge retrieval, utilizing a comprehensive toolbox of 211 tools. The agent evaluates drug interactions at molecular, pharmacokinetic, and clinical levels, identifies contraindications based on patient comorbidities and concurrent medications, and tailors treatment strategies to individual patient characteristics like age, genetic factors, and disease progression. TxAgent retrieves and synthesizes evidence from multiple biomedical sources, assesses drug-patient condition interactions, and refines recommendations through iterative reasoning. It selects tools based on task objectives and executes structured function calls to solve complex therapeutic tasks requiring clinical reasoning and cross-source validation. The ToolUniverse integrates 211 tools, including all US FDA-approved drugs since 1939 and validated clinical insights from Open Targets. TxAgent is open-source and outperforms leading LLMs in drug reasoning and personalized treatment scenarios.

ToolOrchestra is an end-to-end Reinforcement Learning (RL) training framework designed to orchestrate tools and agentic workflows. It allows for the training of small orchestrators that efficiently coordinate the use of intelligent tools, surpassing larger models like GPT-5 in performance while being more efficient. The framework enables the Orchestrator to alternate between reasoning and tool calling in multiple turns, interacting with a diverse set of tools including web search, code interpreters, specialized LLMs, and generalist LLMs. Training is optimized through outcome, efficiency, and preference rewards via end-to-end reinforcement learning, supported by an automatic pipeline for synthesizing environment and tool-call tasks at scale. This framework has been used to produce Orchestrator-8B, a state-of-the-art 8B parameter model for solving complex, multi-turn agentic tasks.

ChainClarity offers AI-powered explanations and deep-dive analysis for over 500 crypto whitepapers, making complex blockchain concepts accessible in plain English. The platform covers a wide range of topics including Bitcoin, Ethereum, DeFi, and NFT projects, eliminating technical jargon. Users can browse trending and new explanations, access detailed breakdowns of tokenomics, risk factors, competitive landscapes, and investment theses. It also features an AI assistant, Qai, for unlimited questions and allows users to create watchlists to track projects. The tool aims to simplify understanding for anyone interested in the crypto space.

HAN offers the source code for implementing and experimenting with Heterogeneous Graph Attention Networks, a model specifically designed for representation learning on heterogeneous graphs. Based on the GAT model, HAN enables researchers and developers to delve into advanced graph neural network architectures. The tool supports various tasks such as node classification and link prediction, making it valuable for academic research in graph theory and machine learning. It provides preprocessed datasets for DBLP, ACM, and IMDB, facilitating quick setup and experimentation. The project also includes detailed documentation on data preprocessing and running experiments, making it accessible for those working with complex graph structures.

KnowledgeEditingPapers is a valuable open-source resource for researchers and practitioners interested in the field of knowledge editing for Large Language Models (LLMs). This GitHub repository compiles a meticulously curated list of academic papers, tutorials, and surveys, offering a comprehensive overview of the latest advancements and methodologies. Users can explore different knowledge editing techniques, including preserve parameters, memory-based, and modify parameters approaches. The repository also highlights recent news, accepted papers, and provides links to relevant resources like Google Drive and BaiduPan for tutorials and reports. It serves as an essential hub for staying updated on the evolving landscape of LLM knowledge editing.

thisrepositorydoesnotexist is an open-source curated list of projects that leverage Machine Learning for synthetic content generation. It serves as a valuable resource for AI researchers and developers looking to explore the landscape of generative AI. The repository categorizes projects by content type, including synthetic websites, images (faces, cats, birds, anime, shoes), and text (lyrics, quotes, resumes, words, game reviews). It also highlights open-source projects like Nvidia Research's Progressive Growing of GANs and OpenAI's Language Models, offering insights into the underlying technologies. This collection is ideal for understanding the breadth of synthetic content applications.

Lowpoly World is an AI tool hosted on Hugging Face Spaces by MirageML, designed for generating images in a lowpoly style. Lowpoly is a distinctive rendering technique that utilizes simplified geometric shapes to create a stylized aesthetic. While the tool aims to provide text-to-image generation capabilities, the current live website content indicates that the application is experiencing a runtime error, preventing its full functionality. It is built using Gradio and is intended to be a free-to-use community-made ML app. Users interested in lowpoly art generation can monitor its status on Hugging Face for resolution of the technical issues.

LoRA + SD Training is an open-source tool available on Hugging Face designed for training AI models, specifically utilizing Low-Rank Adaptation (LoRA) and Stable Diffusion techniques. While the application itself encountered a build error at the time of scraping, the underlying platform, Hugging Face, offers extensive resources for AI development, including various hardware options for Spaces and Inference Endpoints. Users can leverage Hugging Face's infrastructure, which includes a range of CPU and GPU instances, to train and deploy their models. The platform also provides features like private storage, enhanced inference credits, and ZeroGPU quota for subscribers, making it suitable for both personal projects and team-based AI development.

LoRACaptioner is an AI-powered tool designed to generate descriptive captions for images. Users can upload single images or process them in batches, with the option to categorize images for more consistent captioning results. The application allows for the download of images along with their generated captions in a convenient zip file. This feature is particularly useful for content creators, researchers, and anyone needing to quickly add metadata or descriptions to a large collection of images. The tool aims to streamline the process of image annotation and organization.

Legalspace.ai is an AI-first solution designed to revolutionize legal practice for professionals in India. It offers a comprehensive suite of tools including AI-powered legal research, intelligent legal document drafting, and robust case management with annotation capabilities. The platform provides offline access via tablet and iPad apps, ensuring accessibility even without internet connectivity. Legalspace.ai also boasts a vast library of Indian laws and case judgments, matching the size and update frequency of premium subscriptions. Its custom machine learning model assists with Indian legal document drafting, saving significant time and resources. The platform prioritizes security, using its own trained AI models and secure servers to protect user data.