[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-guillaume-chevalier--Awesome-Deep-Learning-Resources":3,"tool-guillaume-chevalier--Awesome-Deep-Learning-Resources":64},[4,17,27,35,43,56],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":16},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,3,"2026-04-05T11:01:52",[13,14,15],"开发框架","图像","Agent","ready",{"id":18,"name":19,"github_repo":20,"description_zh":21,"stars":22,"difficulty_score":23,"last_commit_at":24,"category_tags":25,"status":16},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",140436,2,"2026-04-05T23:32:43",[13,15,26],"语言模型",{"id":28,"name":29,"github_repo":30,"description_zh":31,"stars":32,"difficulty_score":23,"last_commit_at":33,"category_tags":34,"status":16},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",107662,"2026-04-03T11:11:01",[13,14,15],{"id":36,"name":37,"github_repo":38,"description_zh":39,"stars":40,"difficulty_score":23,"last_commit_at":41,"category_tags":42,"status":16},3704,"NextChat","ChatGPTNextWeb\u002FNextChat","NextChat 是一款轻量且极速的 AI 助手，旨在为用户提供流畅、跨平台的大模型交互体验。它完美解决了用户在多设备间切换时难以保持对话连续性，以及面对众多 AI 模型不知如何统一管理的痛点。无论是日常办公、学习辅助还是创意激发，NextChat 都能让用户随时随地通过网页、iOS、Android、Windows、MacOS 或 Linux 端无缝接入智能服务。\n\n这款工具非常适合普通用户、学生、职场人士以及需要私有化部署的企业团队使用。对于开发者而言，它也提供了便捷的自托管方案，支持一键部署到 Vercel 或 Zeabur 等平台。\n\nNextChat 的核心亮点在于其广泛的模型兼容性，原生支持 Claude、DeepSeek、GPT-4 及 Gemini Pro 等主流大模型，让用户在一个界面即可自由切换不同 AI 能力。此外，它还率先支持 MCP（Model Context Protocol）协议，增强了上下文处理能力。针对企业用户，NextChat 提供专业版解决方案，具备品牌定制、细粒度权限控制、内部知识库整合及安全审计等功能，满足公司对数据隐私和个性化管理的高标准要求。",87618,"2026-04-05T07:20:52",[13,26],{"id":44,"name":45,"github_repo":46,"description_zh":47,"stars":48,"difficulty_score":23,"last_commit_at":49,"category_tags":50,"status":16},2268,"ML-For-Beginners","microsoft\u002FML-For-Beginners","ML-For-Beginners 是由微软推出的一套系统化机器学习入门课程，旨在帮助零基础用户轻松掌握经典机器学习知识。这套课程将学习路径规划为 12 周，包含 26 节精炼课程和 52 道配套测验，内容涵盖从基础概念到实际应用的完整流程，有效解决了初学者面对庞大知识体系时无从下手、缺乏结构化指导的痛点。\n\n无论是希望转型的开发者、需要补充算法背景的研究人员，还是对人工智能充满好奇的普通爱好者，都能从中受益。课程不仅提供了清晰的理论讲解，还强调动手实践，让用户在循序渐进中建立扎实的技能基础。其独特的亮点在于强大的多语言支持，通过自动化机制提供了包括简体中文在内的 50 多种语言版本，极大地降低了全球不同背景用户的学习门槛。此外，项目采用开源协作模式，社区活跃且内容持续更新，确保学习者能获取前沿且准确的技术资讯。如果你正寻找一条清晰、友好且专业的机器学习入门之路，ML-For-Beginners 将是理想的起点。",84991,"2026-04-05T10:45:23",[14,51,52,53,15,54,26,13,55],"数据工具","视频","插件","其他","音频",{"id":57,"name":58,"github_repo":59,"description_zh":60,"stars":61,"difficulty_score":10,"last_commit_at":62,"category_tags":63,"status":16},3128,"ragflow","infiniflow\u002Fragflow","RAGFlow 是一款领先的开源检索增强生成（RAG）引擎，旨在为大语言模型构建更精准、可靠的上下文层。它巧妙地将前沿的 RAG 技术与智能体（Agent）能力相结合，不仅支持从各类文档中高效提取知识，还能让模型基于这些知识进行逻辑推理和任务执行。\n\n在大模型应用中，幻觉问题和知识滞后是常见痛点。RAGFlow 通过深度解析复杂文档结构（如表格、图表及混合排版），显著提升了信息检索的准确度，从而有效减少模型“胡编乱造”的现象，确保回答既有据可依又具备时效性。其内置的智能体机制更进一步，使系统不仅能回答问题，还能自主规划步骤解决复杂问题。\n\n这款工具特别适合开发者、企业技术团队以及 AI 研究人员使用。无论是希望快速搭建私有知识库问答系统，还是致力于探索大模型在垂直领域落地的创新者，都能从中受益。RAGFlow 提供了可视化的工作流编排界面和灵活的 API 接口，既降低了非算法背景用户的上手门槛，也满足了专业开发者对系统深度定制的需求。作为基于 Apache 2.0 协议开源的项目，它正成为连接通用大模型与行业专有知识之间的重要桥梁。",77062,"2026-04-04T04:44:48",[15,14,13,26,54],{"id":65,"github_repo":66,"name":67,"description_en":68,"description_zh":69,"ai_summary_zh":69,"readme_en":70,"readme_zh":71,"quickstart_zh":72,"use_case_zh":73,"hero_image_url":74,"owner_login":75,"owner_name":76,"owner_avatar_url":77,"owner_bio":78,"owner_company":79,"owner_location":80,"owner_email":81,"owner_twitter":79,"owner_website":79,"owner_url":82,"languages":79,"stars":83,"forks":84,"last_commit_at":85,"license":86,"difficulty_score":87,"env_os":88,"env_gpu":89,"env_ram":90,"env_deps":91,"category_tags":94,"github_topics":95,"view_count":10,"oss_zip_url":79,"oss_zip_packed_at":79,"status":16,"created_at":103,"updated_at":104,"faqs":105,"releases":106},470,"guillaume-chevalier\u002FAwesome-Deep-Learning-Resources","Awesome-Deep-Learning-Resources","Rough list of my favorite deep learning resources, useful for revisiting topics or for reference. I have got through all of the content listed there, carefully. - Guillaume Chevalier","Awesome-Deep-Learning-Resources 是由开发者 Guillaume Chevalier 主导的深度学习资源整理项目，旨在为学习者和从业者提供系统化的知识地图。它通过精心分类的 8 大模块（包括在线课程、书籍、论文、实践工具等），覆盖从基础概念到前沿技术的完整学习路径，尤其注重理论与实践结合。项目特别收录了作者亲自验证过的课程、代码库和数据集，并附带 Google 趋势分析图，帮助用户把握领域发展脉络。\n\n该项目解决了深度学习学习资源分散、缺乏系统性指导的问题，尤其适合需要快速建立知识框架的开发者、研究人员和高校学生。其独特价值在于：1）资源经过作者逐项验证，避免无效信息；2）包含罕见的数学理论专题（如梯度下降算法、复数信号处理）；3）前瞻性地探讨 GPU 架构演进与量子计算对深度学习的影响。对于希望从零基础入门或需要持续跟踪技术动态的用户，这里提供了从 Andrew Ng 经典课程到最新论文的完整生态链，同时推荐的实战工具和数据集能直接应用于项目开发。","# [Awesome Deep Learning Resources](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FAwesome-Deep-Learning-Resources) [![Awesome](https:\u002F\u002Fcdn.rawgit.com\u002Fsindresorhus\u002Fawesome\u002Fd7305f38d29fed78fa85652e3a63e154dd8e8829\u002Fmedia\u002Fbadge.svg)](https:\u002F\u002Fgithub.com\u002Fsindresorhus\u002Fawesome)\n\nThis is a rough list of my favorite deep learning resources. It has been useful to me for learning how to do deep learning, I use it for revisiting topics or for reference.\nI ([Guillaume Chevalier](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier)) have built this list and got through all of the content listed here, carefully.\n\n\n## Contents\n\n- [Trends](#trends)\n- [Online classes](#online-classes)\n- [Books](#books)\n- [Posts and Articles](#posts-and-articles)\n- [Practical resources](#practical-resources)\n  - [Librairies and Implementations](#librairies-and-implementations)\n  - [Some Datasets](#some-datasets)\n- [Other Math Theory](#other-math-theory)\n  - [Gradient Descent Algorithms and optimization](#gradient-descent-algorithms-and-optimization)\n  - [Complex Numbers & Digital Signal Processing](#complex-numbers-and-digital-signal-processing)\n- [Papers](#papers)\n  - [Recurrent Neural Networks](#recurrent-neural-networks)\n  - [Convolutional Neural Networks](#convolutional-neural-networks)\n  - [Attention Mechanisms](#attention-mechanisms)\n  - [Other](#other)\n- [YouTube and Videos](#youtube)\n- [Misc. Hubs and Links](#misc-hubs-and-links)\n- [License](#license)\n\n\u003Ca name=\"trends\" \u002F>\n\n## Trends\n\nHere are the all-time [Google Trends](https:\u002F\u002Fwww.google.ca\u002Ftrends\u002Fexplore?date=all&q=machine%20learning,deep%20learning,data%20science,computer%20programming), from 2004 up to now, September 2017:\n\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fguillaume-chevalier_Awesome-Deep-Learning-Resources_readme_940d3fb12df2.png\" width=\"792\" height=\"424\" \u002F>\n\u003C\u002Fp>\n\nYou might also want to look at Andrej Karpathy's [new post](https:\u002F\u002Fmedium.com\u002F@karpathy\u002Fa-peek-at-trends-in-machine-learning-ab8a1085a106) about trends in Machine Learning research.\n\nI believe that Deep learning is the key to make computers think more like humans, and has a lot of potential. Some hard automation tasks can be solved easily with that while this was impossible to achieve earlier with classical algorithms.\n\nMoore's Law about exponential progress rates in computer science hardware is now more affecting GPUs than CPUs because of physical limits on how tiny an atomic transistor can be. We are shifting toward parallel architectures\n[[read more](https:\u002F\u002Fwww.quora.com\u002FDoes-Moores-law-apply-to-GPUs-Or-only-CPUs)]. Deep learning exploits parallel architectures as such under the hood by using GPUs. On top of that, deep learning algorithms may use Quantum Computing and apply to machine-brain interfaces in the future.\n\nI find that the key of intelligence and cognition is a very interesting subject to explore and is not yet well understood. Those technologies are promising.\n\n\n\u003Ca name=\"online-classes\" \u002F>\n\n## Online Classes\n\n- **[DL&RNN Course](https:\u002F\u002Fwww.dl-rnn-course.neuraxio.com\u002Fstart?utm_source=github_awesome) - I created this richely dense course on Deep Learning and Recurrent Neural Networks.**\n- [Machine Learning by Andrew Ng on Coursera](https:\u002F\u002Fwww.coursera.org\u002Flearn\u002Fmachine-learning) - Renown entry-level online class with [certificate](https:\u002F\u002Fwww.coursera.org\u002Faccount\u002Faccomplishments\u002Fverify\u002FDXPXHYFNGKG3). Taught by: Andrew Ng, Associate Professor, Stanford University; Chief Scientist, Baidu; Chairman and Co-founder, Coursera.\n- [Deep Learning Specialization by Andrew Ng on Coursera](https:\u002F\u002Fwww.coursera.org\u002Fspecializations\u002Fdeep-learning) - New series of 5 Deep Learning courses by Andrew Ng, now with Python rather than Matlab\u002FOctave, and which leads to a [specialization certificate](https:\u002F\u002Fwww.coursera.org\u002Faccount\u002Faccomplishments\u002Fspecialization\u002FU7VNC3ZD9YD8).\n- [Deep Learning by Google](https:\u002F\u002Fwww.udacity.com\u002Fcourse\u002Fdeep-learning--ud730) - Good intermediate to advanced-level course covering high-level deep learning concepts, I found it helps to get creative once the basics are acquired.\n- [Machine Learning for Trading by Georgia Tech](https:\u002F\u002Fwww.udacity.com\u002Fcourse\u002Fmachine-learning-for-trading--ud501) - Interesting class for acquiring basic knowledge of machine learning applied to trading and some AI and finance concepts. I especially liked the section on Q-Learning.\n- [Neural networks class by Hugo Larochelle, Université de Sherbrooke](https:\u002F\u002Fwww.youtube.com\u002Fplaylist?list=PL6Xpj9I5qXYEcOhn7TqghAJ6NAPrNmUBH) - Interesting class about neural networks available online for free by Hugo Larochelle, yet I have watched a few of those videos.\n- [GLO-4030\u002F7030 Apprentissage par réseaux de neurones profonds](https:\u002F\u002Fulaval-damas.github.io\u002Fglo4030\u002F) - This is a class given by Philippe Giguère, Professor at University Laval. I especially found awesome its rare visualization of the multi-head attention mechanism, which can be contemplated at the [slide 28 of week 13's class](http:\u002F\u002Fwww2.ift.ulaval.ca\u002F~pgiguere\u002Fcours\u002FDeepLearning\u002F09-Attention.pdf).\n- [Deep Learning & Recurrent Neural Networks (DL&RNN)](https:\u002F\u002Fwww.neuraxio.com\u002Fen\u002Ftime-series-solution) - The most richly dense, accelerated course on the topic of Deep Learning & Recurrent Neural Networks (scroll at the end).\n\n\u003Ca name=\"books\" \u002F>\n\n## Books\n\n- [Clean Code](https:\u002F\u002Fwww.amazon.ca\u002FClean-Code-Handbook-Software-Craftsmanship\u002Fdp\u002F0132350882) - Get back to the basics you fool! Learn how to do Clean Code for your career. This is by far the best book I've read even if this list is related to Deep Learning.\n- [Clean Coder](https:\u002F\u002Fwww.amazon.ca\u002FClean-Coder-Conduct-Professional-Programmers\u002Fdp\u002F0137081073) - Learn how to be professional as a coder and how to interact with your manager. This is important for any coding career.\n- [How to Create a Mind](https:\u002F\u002Fwww.amazon.com\u002FHow-Create-Mind-Thought-Revealed\u002Fdp\u002FB009VSFXZ4) - The audio version is nice to listen to while commuting. This book is motivating about reverse-engineering the mind and thinking on how to code AI.\n- [Neural Networks and Deep Learning](http:\u002F\u002Fneuralnetworksanddeeplearning.com\u002Findex.html) - This book covers many of the core concepts behind neural networks and deep learning.\n- [Deep Learning - An MIT Press book](http:\u002F\u002Fwww.deeplearningbook.org\u002F) - Yet halfway through the book, it contains satisfying math content on how to think about actual deep learning.\n- [Some other books I have read](https:\u002F\u002Fbooks.google.ca\u002Fbooks?hl=en&as_coll=4&num=100&uid=103409002069648430166&source=gbs_slider_cls_metadata_4_mylibrary_title) - Some books listed here are less related to deep learning but are still somehow relevant to this list.\n\n\u003Ca name=\"posts-and-articles\" \u002F>\n\n## Posts and Articles\n\n- [Predictions made by Ray Kurzweil](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FPredictions_made_by_Ray_Kurzweil) - List of mid to long term futuristic predictions made by Ray Kurzweil.\n- [The Unreasonable Effectiveness of Recurrent Neural Networks](http:\u002F\u002Fkarpathy.github.io\u002F2015\u002F05\u002F21\u002Frnn-effectiveness\u002F) - MUST READ post by Andrej Karpathy - this is what motivated me to learn RNNs, it demonstrates what it can achieve in the most basic form of NLP.\n- [Neural Networks, Manifolds, and Topology](http:\u002F\u002Fcolah.github.io\u002Fposts\u002F2014-03-NN-Manifolds-Topology\u002F) - Fresh look on how neurons map information.\n- [Understanding LSTM Networks](http:\u002F\u002Fcolah.github.io\u002Fposts\u002F2015-08-Understanding-LSTMs\u002F) - Explains the LSTM cells' inner workings, plus, it has interesting links in conclusion.\n- [Attention and Augmented Recurrent Neural Networks](http:\u002F\u002Fdistill.pub\u002F2016\u002Faugmented-rnns\u002F) - Interesting for visual animations, it is a nice intro to attention mechanisms as an example.\n- [Recommending music on Spotify with deep learning](http:\u002F\u002Fbenanne.github.io\u002F2014\u002F08\u002F05\u002Fspotify-cnns.html) - Awesome for doing clustering on audio - post by an intern at Spotify.\n- [Announcing SyntaxNet: The World’s Most Accurate Parser Goes Open Source](https:\u002F\u002Fresearch.googleblog.com\u002F2016\u002F05\u002Fannouncing-syntaxnet-worlds-most.html) - Parsey McParseface's birth, a neural syntax tree parser.\n- [Improving Inception and Image Classification in TensorFlow](https:\u002F\u002Fresearch.googleblog.com\u002F2016\u002F08\u002Fimproving-inception-and-image.html) - Very interesting CNN architecture (e.g.: the inception-style convolutional layers is promising and efficient in terms of reducing the number of parameters).\n- [WaveNet: A Generative Model for Raw Audio](https:\u002F\u002Fdeepmind.com\u002Fblog\u002Fwavenet-generative-model-raw-audio\u002F) - Realistic talking machines: perfect voice generation.\n- [François Chollet's Twitter](https:\u002F\u002Ftwitter.com\u002Ffchollet) - Author of Keras - has interesting Twitter posts and innovative ideas.\n- [Neuralink and the Brain’s Magical Future](http:\u002F\u002Fwaitbutwhy.com\u002F2017\u002F04\u002Fneuralink.html) - Thought provoking article about the future of the brain and brain-computer interfaces.\n- [Migrating to Git LFS for Developing Deep Learning Applications with Large Files](http:\u002F\u002Fvooban.com\u002Fen\u002Ftips-articles-geek-stuff\u002Fmigrating-to-git-lfs-for-developing-deep-learning-applications-with-large-files\u002F) - Easily manage huge files in your private Git projects.\n- [The future of deep learning](https:\u002F\u002Fblog.keras.io\u002Fthe-future-of-deep-learning.html) - François Chollet's thoughts on the future of deep learning.\n- [Discover structure behind data with decision trees](http:\u002F\u002Fvooban.com\u002Fen\u002Ftips-articles-geek-stuff\u002Fdiscover-structure-behind-data-with-decision-trees\u002F) - Grow decision trees and visualize them, infer the hidden logic behind data.\n- [Hyperopt tutorial for Optimizing Neural Networks’ Hyperparameters](http:\u002F\u002Fvooban.com\u002Fen\u002Ftips-articles-geek-stuff\u002Fhyperopt-tutorial-for-optimizing-neural-networks-hyperparameters\u002F) - Learn to slay down hyperparameter spaces automatically rather than by hand.\n- [Estimating an Optimal Learning Rate For a Deep Neural Network](https:\u002F\u002Fmedium.com\u002F@surmenok\u002Festimating-optimal-learning-rate-for-a-deep-neural-network-ce32f2556ce0) - Clever trick to estimate an optimal learning rate prior any single full training.\n - [The Annotated Transformer](http:\u002F\u002Fnlp.seas.harvard.edu\u002F2018\u002F04\u002F03\u002Fattention.html) - Good for understanding the \"Attention Is All You Need\" (AIAYN) paper. \n - [The Illustrated Transformer](http:\u002F\u002Fjalammar.github.io\u002Fillustrated-transformer\u002F) - Also good for understanding the \"Attention Is All You Need\" (AIAYN) paper.\n - [Improving Language Understanding with Unsupervised Learning](https:\u002F\u002Fblog.openai.com\u002Flanguage-unsupervised\u002F) - SOTA across many NLP tasks from unsupervised pretraining on huge corpus.\n - [NLP's ImageNet moment has arrived](https:\u002F\u002Fthegradient.pub\u002Fnlp-imagenet\u002F) - All hail NLP's ImageNet moment. \n - [The Illustrated BERT, ELMo, and co. (How NLP Cracked Transfer Learning)](https:\u002F\u002Fjalammar.github.io\u002Fillustrated-bert\u002F) - Understand the different approaches used for NLP's ImageNet moment. \n - [Uncle Bob's Principles Of OOD](http:\u002F\u002Fbutunclebob.com\u002FArticleS.UncleBob.PrinciplesOfOod) - Not only the SOLID principles are needed for doing clean code, but the furtherless known REP, CCP, CRP, ADP, SDP and SAP principles are very important for developping huge software that must be bundled in different separated packages.\n - [Why do 87% of data science projects never make it into production?](https:\u002F\u002Fventurebeat.com\u002F2019\u002F07\u002F19\u002Fwhy-do-87-of-data-science-projects-never-make-it-into-production\u002F) - Data is not to be overlooked, and communication between teams and data scientists is important to integrate solutions properly.\n - [The real reason most ML projects fail](https:\u002F\u002Ftowardsdatascience.com\u002Fwhat-is-the-main-reason-most-ml-projects-fail-515d409a161f) - Focus on clear business objectives, avoid pivots of algorithms unless you have really clean code, and be able to know when what you coded is \"good enough\".\n - [SOLID Machine Learning](https:\u002F\u002Fwww.umaneo.com\u002Fpost\u002Fthe-solid-principles-applied-to-machine-learning) - The SOLID principles applied to Machine Learning.\n \n\u003Ca name=\"practical-resources\" \u002F>\n\n## Practical Resources\n\n\u003Ca name=\"librairies-and-implementations\" \u002F>\n\n### Librairies and Implementations\n- [Neuraxle, a framwework for machine learning pipelines](https:\u002F\u002Fgithub.com\u002FNeuraxio\u002FNeuraxle) - The best framework for structuring and deploying your machine learning projects, and which is also compatible with most framework (e.g.: Scikit-Learn, TensorFlow, PyTorch, Keras, and so forth).\n- [TensorFlow's GitHub repository](https:\u002F\u002Fgithub.com\u002Ftensorflow\u002Ftensorflow) - Most known deep learning framework, both high-level and low-level while staying flexible.\n- [skflow](https:\u002F\u002Fgithub.com\u002Ftensorflow\u002Fskflow) - TensorFlow wrapper à la scikit-learn.\n- [Keras](https:\u002F\u002Fkeras.io\u002F) - Keras is another intersting deep learning framework like TensorFlow, it is mostly high-level.\n- [carpedm20's repositories](https:\u002F\u002Fgithub.com\u002Fcarpedm20) - Many interesting neural network architectures are implemented by the Korean guy Taehoon Kim, A.K.A. carpedm20.\n- [carpedm20\u002FNTM-tensorflow](https:\u002F\u002Fgithub.com\u002Fcarpedm20\u002FNTM-tensorflow) - Neural Turing Machine TensorFlow implementation.\n- [Deep learning for lazybones](http:\u002F\u002Foduerr.github.io\u002Fblog\u002F2016\u002F04\u002F06\u002FDeep-Learning_for_lazybones) - Transfer learning tutorial in TensorFlow for vision from high-level embeddings of a pretrained CNN, AlexNet 2012.\n- [LSTM for Human Activity Recognition (HAR)](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FLSTM-Human-Activity-Recognition) - Tutorial of mine on using LSTMs on time series for classification.\n- [Deep stacked residual bidirectional LSTMs for HAR](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FHAR-stacked-residual-bidir-LSTMs) - Improvements on the previous project.\n- [Sequence to Sequence (seq2seq) Recurrent Neural Network (RNN) for Time Series Prediction](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002Fseq2seq-signal-prediction) - Tutorial of mine on how to predict temporal sequences of numbers - that may be multichannel.\n- [Hyperopt for a Keras CNN on CIFAR-100](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FHyperopt-Keras-CNN-CIFAR-100) - Auto (meta) optimizing a neural net (and its architecture) on the CIFAR-100 dataset.\n- [ML \u002F DL repositories I starred](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier?direction=desc&page=1&q=machine+OR+deep+OR+learning+OR+rnn+OR+lstm+OR+cnn&sort=stars&tab=stars&utf8=%E2%9C%93) - GitHub is full of nice code samples & projects.\n- [Smoothly Blend Image Patches](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FSmoothly-Blend-Image-Patches) - Smooth patch merger for [semantic segmentation with a U-Net](https:\u002F\u002Fvooban.com\u002Fen\u002Ftips-articles-geek-stuff\u002Fsatellite-image-segmentation-workflow-with-u-net\u002F).\n- [Self Governing Neural Networks (SGNN): the Projection Layer](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FSGNN-Self-Governing-Neural-Networks-Projection-Layer) - With this, you can use words in your deep learning models without training nor loading embeddings.\n- [Neuraxle](https:\u002F\u002Fgithub.com\u002FNeuraxio\u002FNeuraxle) - Neuraxle is a Machine Learning (ML) library for building neat pipelines, providing the right abstractions to both ease research, development, and deployment of your ML applications.\n- [Clean Machine Learning, a Coding Kata](https:\u002F\u002Fgithub.com\u002FNeuraxio\u002FKata-Clean-Machine-Learning-From-Dirty-Code) - Learn the good design patterns to use for doing Machine Learning the good way, by practicing.\n\n\u003Ca name=\"some-datasets\" \u002F>\n\n### Some Datasets\n\nThose are resources I have found that seems interesting to develop models onto.\n\n- [UCI Machine Learning Repository](https:\u002F\u002Farchive.ics.uci.edu\u002Fml\u002Fdatasets.html) - TONS of datasets for ML.\n- [Cornell Movie--Dialogs Corpus](http:\u002F\u002Fwww.cs.cornell.edu\u002F~cristian\u002FCornell_Movie-Dialogs_Corpus.html) - This could be used for a chatbot.\n- [SQuAD The Stanford Question Answering Dataset](https:\u002F\u002Frajpurkar.github.io\u002FSQuAD-explorer\u002F) - Question answering dataset that can be explored online, and a list of models performing well on that dataset.\n- [LibriSpeech ASR corpus](http:\u002F\u002Fwww.openslr.org\u002F12\u002F) - Huge free English speech dataset with balanced genders and speakers, that seems to be of high quality.\n- [Awesome Public Datasets](https:\u002F\u002Fgithub.com\u002Fcaesar0301\u002Fawesome-public-datasets) - An awesome list of public datasets.\n- [SentEval: An Evaluation Toolkit for Universal Sentence Representations](https:\u002F\u002Farxiv.org\u002Fabs\u002F1803.05449) - A Python framework to benchmark your sentence representations on many datasets (NLP tasks). \n- [ParlAI: A Dialog Research Software Platform](https:\u002F\u002Farxiv.org\u002Fabs\u002F1705.06476) - Another Python framework to benchmark your sentence representations on many datasets (NLP tasks).\n\n\n\u003Ca name=\"other-math-theory\" \u002F>\n\n## Other Math Theory\n\n\u003Ca name=\"gradient-descent-algorithms-and-optimization\" \u002F>\n\n### Gradient Descent Algorithms & Optimization Theory\n\n- [Neural Networks and Deep Learning, ch.2](http:\u002F\u002Fneuralnetworksanddeeplearning.com\u002Fchap2.html) - Overview on how does the backpropagation algorithm works.\n- [Neural Networks and Deep Learning, ch.4](http:\u002F\u002Fneuralnetworksanddeeplearning.com\u002Fchap4.html) - A visual proof that neural nets can compute any function.\n- [Yes you should understand backprop](https:\u002F\u002Fmedium.com\u002F@karpathy\u002Fyes-you-should-understand-backprop-e2f06eab496b#.mr5wq61fb) - Exposing backprop's caveats and the importance of knowing that while training models.\n- [Artificial Neural Networks: Mathematics of Backpropagation](http:\u002F\u002Fbriandolhansky.com\u002Fblog\u002F2013\u002F9\u002F27\u002Fartificial-neural-networks-backpropagation-part-4) - Picturing backprop, mathematically.\n- [Deep Learning Lecture 12: Recurrent Neural Nets and LSTMs](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=56TYLaQN4N8) - Unfolding of RNN graphs is explained properly, and potential problems about gradient descent algorithms are exposed.\n- [Gradient descent algorithms in a saddle point](http:\u002F\u002Fsebastianruder.com\u002Fcontent\u002Fimages\u002F2016\u002F09\u002Fsaddle_point_evaluation_optimizers.gif) - Visualize how different optimizers interacts with a saddle points.\n- [Gradient descent algorithms in an almost flat landscape](https:\u002F\u002Fdevblogs.nvidia.com\u002Fwp-content\u002Fuploads\u002F2015\u002F12\u002FNKsFHJb.gif) - Visualize how different optimizers interacts with an almost flat landscape.\n- [Gradient Descent](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=F6GSRDoB-Cg) - Okay, I already listed Andrew NG's Coursera class above, but this video especially is quite pertinent as an introduction and defines the gradient descent algorithm.\n- [Gradient Descent: Intuition](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=YovTqTY-PYY) - What follows from the previous video: now add intuition.\n- [Gradient Descent in Practice 2: Learning Rate](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=gX6fZHgfrow) - How to adjust the learning rate of a neural network.\n- [The Problem of Overfitting](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=u73PU6Qwl1I) - A good explanation of overfitting and how to address that problem.\n- [Diagnosing Bias vs Variance](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=ewogYw5oCAI) - Understanding bias and variance in the predictions of a neural net and how to address those problems.\n- [Self-Normalizing Neural Networks](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1706.02515.pdf) - Appearance of the incredible SELU activation function.\n- [Learning to learn by gradient descent by gradient descent](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1606.04474.pdf) - RNN as an optimizer: introducing the L2L optimizer, a meta-neural network.\n\n\u003Ca name=\"complex-numbers-and-digital-signal-processing\" \u002F>\n\n### Complex Numbers & Digital Signal Processing\n\nOkay, signal processing might not be directly related to deep learning, but studying it is interesting to have more intuition in developing neural architectures based on signal.\n\n- [Window Functions](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FWindow_function) - Wikipedia page that lists some of the known window functions - note that the [Hann-Poisson window](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FWindow_function#Hann%E2%80%93Poisson_window) is specially interesting for greedy hill-climbing algorithms (like gradient descent for example). \n- [MathBox, Tools for Thought Graphical Algebra and Fourier Analysis](https:\u002F\u002Facko.net\u002Ffiles\u002Fgltalks\u002Ftoolsforthought\u002F) - New look on Fourier analysis.\n- [How to Fold a Julia Fractal](http:\u002F\u002Facko.net\u002Fblog\u002Fhow-to-fold-a-julia-fractal\u002F) - Animations dealing with complex numbers and wave equations.\n- [Animate Your Way to Glory, Math and Physics in Motion](http:\u002F\u002Facko.net\u002Fblog\u002Fanimate-your-way-to-glory\u002F) - Convergence methods in physic engines, and applied to interaction design.\n- [Animate Your Way to Glory - Part II, Math and Physics in Motion](http:\u002F\u002Facko.net\u002Fblog\u002Fanimate-your-way-to-glory-pt2\u002F) - Nice animations for rotation and rotation interpolation with Quaternions, a mathematical object for handling 3D rotations.\n- [Filtering signal, plotting the STFT and the Laplace transform](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002Ffiltering-stft-and-laplace-transform) - Simple Python demo on signal processing.\n\n\n\u003Ca name=\"papers\" \u002F>\n\n## Papers\n\n\u003Ca name=\"recurrent-neural-networks\" \u002F>\n\n### Recurrent Neural Networks\n\n- [Deep Learning in Neural Networks: An Overview](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1404.7828v4.pdf) - You_Again's summary\u002Foverview of deep learning, mostly about RNNs.\n- [Bidirectional Recurrent Neural Networks](http:\u002F\u002Fwww.di.ufpe.br\u002F~fnj\u002FRNA\u002Fbibliografia\u002FBRNN.pdf) - Better classifications with RNNs with bidirectional scanning on the time axis.\n- [Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1406.1078v3.pdf) - Two networks in one combined into a seq2seq (sequence to sequence) Encoder-Decoder architecture. RNN Encoder–Decoder with 1000 hidden units. Adadelta optimizer.\n- [Sequence to Sequence Learning with Neural Networks](http:\u002F\u002Fpapers.nips.cc\u002Fpaper\u002F5346-sequence-to-sequence-learning-with-neural-networks.pdf) - 4 stacked LSTM cells of 1000 hidden size with reversed input sentences, and with beam search, on the WMT’14 English to French dataset.\n- [Exploring the Limits of Language Modeling](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1602.02410.pdf) - Nice recursive models using word-level LSTMs on top of a character-level CNN using an overkill amount of GPU power.\n- [Neural Machine Translation and Sequence-to-sequence Models: A Tutorial](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1703.01619.pdf) - Interesting overview of the subject of NMT, I mostly read part 8 about RNNs with attention as a refresher.\n- [Exploring the Depths of Recurrent Neural Networks with Stochastic Residual Learning](https:\u002F\u002Fcs224d.stanford.edu\u002Freports\u002FPradhanLongpre.pdf) - Basically, residual connections can be better than stacked RNNs in the presented case of sentiment analysis.\n- [Pixel Recurrent Neural Networks](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1601.06759.pdf) - Nice for photoshop-like \"content aware fill\" to fill missing patches in images.\n- [Adaptive Computation Time for Recurrent Neural Networks](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1603.08983v4.pdf) - Let RNNs decide how long they compute. I would love to see how well would it combines to Neural Turing Machines. Interesting interactive visualizations on the subject can be found [here](http:\u002F\u002Fdistill.pub\u002F2016\u002Faugmented-rnns\u002F).\n\n\u003Ca name=\"convolutional-neural-networks\" \u002F>\n\n### Convolutional Neural Networks\n\n- [What is the Best Multi-Stage Architecture for Object Recognition?](http:\u002F\u002Fyann.lecun.com\u002Fexdb\u002Fpublis\u002Fpdf\u002Fjarrett-iccv-09.pdf) - Awesome for the use of \"local contrast normalization\".\n- [ImageNet Classification with Deep Convolutional Neural Networks](http:\u002F\u002Fwww.cs.toronto.edu\u002F~fritz\u002Fabsps\u002Fimagenet.pdf) - AlexNet, 2012 ILSVRC, breakthrough of the ReLU activation function.\n- [Visualizing and Understanding Convolutional Networks](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1311.2901v3.pdf) - For the \"deconvnet layer\".\n- [Fast and Accurate Deep Network Learning by Exponential Linear Units](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1511.07289v1.pdf) - ELU activation function for CIFAR vision tasks.\n- [Very Deep Convolutional Networks for Large-Scale Image Recognition](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1409.1556v6.pdf) - Interesting idea of stacking multiple 3x3 conv+ReLU before pooling for a bigger filter size with just a few parameters. There is also a nice table for \"ConvNet Configuration\".\n- [Going Deeper with Convolutions](http:\u002F\u002Fwww.cv-foundation.org\u002Fopenaccess\u002Fcontent_cvpr_2015\u002Fpapers\u002FSzegedy_Going_Deeper_With_2015_CVPR_paper.pdf) - GoogLeNet: Appearance of \"Inception\" layers\u002Fmodules, the idea is of parallelizing conv layers into many mini-conv of different size with \"same\" padding, concatenated on depth.\n- [Highway Networks](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1505.00387v2.pdf) - Highway networks: residual connections.\n- [Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1502.03167v3.pdf) - Batch normalization (BN): to normalize a layer's output by also summing over the entire batch, and then performing a linear rescaling and shifting of a certain trainable amount.\n- [U-Net: Convolutional Networks for Biomedical Image Segmentation](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1505.04597.pdf) - The U-Net is an encoder-decoder CNN that also has skip-connections, good for image segmentation at a per-pixel level.\n- [Deep Residual Learning for Image Recognition](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1512.03385v1.pdf) - Very deep residual layers with batch normalization layers - a.k.a. \"how to overfit any vision dataset with too many layers and make any vision model work properly at recognition given enough data\".\n- [Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1602.07261v2.pdf) - For improving GoogLeNet with residual connections.\n- [WaveNet: a Generative Model for Raw Audio](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1609.03499v2.pdf) - Epic raw voice\u002Fmusic generation with new architectures based on dilated causal convolutions to capture more audio length.\n- [Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1610.07584v2.pdf) - 3D-GANs for 3D model generation and fun 3D furniture arithmetics from embeddings (think like word2vec word arithmetics with 3D furniture representations).\n- [Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour](https:\u002F\u002Fresearch.fb.com\u002Fpublications\u002FImageNet1kIn1h\u002F) - Incredibly fast distributed training of a CNN.\n- [Densely Connected Convolutional Networks](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1608.06993.pdf) - Best Paper Award at CVPR 2017, yielding improvements on state-of-the-art performances on CIFAR-10, CIFAR-100 and SVHN datasets, this new neural network architecture is named DenseNet.\n- [The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1611.09326.pdf) - Merges the ideas of the U-Net and the DenseNet, this new neural network is especially good for huge datasets in image segmentation.\n- [Prototypical Networks for Few-shot Learning](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1703.05175.pdf) - Use a distance metric in the loss to determine to which class does an object belongs to from a few examples.\n\n\u003Ca name=\"attention-mechanisms\" \u002F>\n\n### Attention Mechanisms\n\n- [Neural Machine Translation by Jointly Learning to Align and Translate](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1409.0473.pdf) - Attention mechanism for LSTMs! Mostly, figures and formulas and their explanations revealed to be useful to me. I gave a talk on that paper [here](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QuvRWevJMZ4).\n- [Neural Turing Machines](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1410.5401v2.pdf) - Outstanding for letting a neural network learn an algorithm with seemingly good generalization over long time dependencies. Sequences recall problem.\n- [Show, Attend and Tell: Neural Image Caption Generation with Visual Attention](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1502.03044.pdf) - LSTMs' attention mechanisms on CNNs feature maps does wonders.\n- [Teaching Machines to Read and Comprehend](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1506.03340v3.pdf) - A very interesting and creative work about textual question answering, what a breakthrough, there is something to do with that.\n- [Effective Approaches to Attention-based Neural Machine Translation](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1508.04025.pdf) - Exploring different approaches to attention mechanisms.\n- [Matching Networks for One Shot Learning](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1606.04080.pdf) - Interesting way of doing one-shot learning with low-data by using an attention mechanism and a query to compare an image to other images for classification.\n- [Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1609.08144.pdf) - In 2016: stacked residual LSTMs with attention mechanisms on encoder\u002Fdecoder are the best for NMT (Neural Machine Translation).\n- [Hybrid computing using a neural network with dynamic external memory](http:\u002F\u002Fwww.nature.com\u002Farticles\u002Fnature20101.epdf?author_access_token=ImTXBI8aWbYxYQ51Plys8NRgN0jAjWel9jnR3ZoTv0MggmpDmwljGswxVdeocYSurJ3hxupzWuRNeGvvXnoO8o4jTJcnAyhGuZzXJ1GEaD-Z7E6X_a9R-xqJ9TfJWBqz) - Improvements on differentiable memory based on NTMs: now it is the Differentiable Neural Computer (DNC).\n- [Massive Exploration of Neural Machine Translation Architectures](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1703.03906.pdf) - That yields intuition about the boundaries of what works for doing NMT within a framed seq2seq problem formulation.\n- [Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram\nPredictions](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1712.05884.pdf) - A [WaveNet](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1609.03499v2.pdf) used as a vocoder can be conditioned on generated Mel Spectrograms from the Tacotron 2 LSTM neural network with attention to generate neat audio from text.\n- [Attention Is All You Need](https:\u002F\u002Farxiv.org\u002Fabs\u002F1706.03762) (AIAYN) - Introducing multi-head self-attention neural networks with positional encoding to do sentence-level NLP without any RNN nor CNN - this paper is a must-read (also see [this explanation](http:\u002F\u002Fnlp.seas.harvard.edu\u002F2018\u002F04\u002F03\u002Fattention.html) and [this visualization](http:\u002F\u002Fjalammar.github.io\u002Fillustrated-transformer\u002F) of the paper). \n\n\u003Ca name=\"other\" \u002F>\n\n### Other\n\n- [ProjectionNet: Learning Efficient On-Device Deep Networks Using Neural Projections](https:\u002F\u002Farxiv.org\u002Fabs\u002F1708.00630) - Replace word embeddings by word projections in your deep neural networks, which doesn't require a pre-extracted dictionnary nor storing embedding matrices. \n- [Self-Governing Neural Networks for On-Device Short Text Classification](http:\u002F\u002Faclweb.org\u002Fanthology\u002FD18-1105) - This paper is the sequel to the ProjectionNet just above. The SGNN is elaborated on the ProjectionNet, and the optimizations are detailed more in-depth (also see my [attempt to reproduce the paper in code](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FSGNN-Self-Governing-Neural-Networks-Projection-Layer) and watch [the talks' recording](https:\u002F\u002Fvimeo.com\u002F305197775)).\n- [Matching Networks for One Shot Learning](https:\u002F\u002Farxiv.org\u002Fabs\u002F1606.04080) - Classify a new example from a list of other examples (without definitive categories) and with low-data per classification task, but lots of data for lots of similar classification tasks - it seems better than siamese networks. To sum up: with Matching Networks, you can optimize directly for a cosine similarity between examples (like a self-attention product would match) which is passed to the softmax directly. I guess that Matching Networks could probably be used as with negative-sampling softmax training in word2vec's CBOW or Skip-gram without having to do any context embedding lookups. \n\n\n\u003Ca name=\"youtube\" \u002F>\n\n## YouTube and Videos\n\n- [Attention Mechanisms in Recurrent Neural Networks (RNNs) - IGGG](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QuvRWevJMZ4) - A talk for a reading group on attention mechanisms (Paper: Neural Machine Translation by Jointly Learning to Align and Translate).\n- [Tensor Calculus and the Calculus of Moving Surfaces](https:\u002F\u002Fwww.youtube.com\u002Fplaylist?list=PLlXfTHzgMRULkodlIEqfgTS-H1AY_bNtq) - Generalize properly how Tensors work, yet just watching a few videos already helps a lot to grasp the concepts.\n- [Deep Learning & Machine Learning (Advanced topics)](https:\u002F\u002Fwww.youtube.com\u002Fplaylist?list=PLlp-GWNOd6m4C_-9HxuHg2_ZeI2Yzwwqt) - A list of videos about deep learning that I found interesting or useful, this is a mix of a bit of everything.\n- [Signal Processing Playlist](https:\u002F\u002Fwww.youtube.com\u002Fplaylist?list=PLlp-GWNOd6m6gSz0wIcpvl4ixSlS-HEmr) - A YouTube playlist I composed about DFT\u002FFFT, STFT and the Laplace transform - I was mad about my software engineering bachelor not including signal processing classes (except a bit in the quantum physics class).\n- [Computer Science](https:\u002F\u002Fwww.youtube.com\u002Fplaylist?list=PLlp-GWNOd6m7vLOsW20xAJ81-65C-Ys6k) - Yet another YouTube playlist I composed, this time about various CS topics.\n- [Siraj's Channel](https:\u002F\u002Fwww.youtube.com\u002Fchannel\u002FUCWN3xxRkmTPmbKwht9FuE5A\u002Fvideos?view=0&sort=p&flow=grid) - Siraj has entertaining, fast-paced video tutorials about deep learning.\n- [Two Minute Papers' Channel](https:\u002F\u002Fwww.youtube.com\u002Fuser\u002Fkeeroyz\u002Fvideos?sort=p&view=0&flow=grid) - Interesting and shallow overview of some research papers, for example about WaveNet or Neural Style Transfer.\n- [Geoffrey Hinton interview](https:\u002F\u002Fwww.coursera.org\u002Flearn\u002Fneural-networks-deep-learning\u002Flecture\u002Fdcm5r\u002Fgeoffrey-hinton-interview) - Andrew Ng interviews Geoffrey Hinton, who talks about his research and breaktroughs, and gives advice for students.\n- [Growing Neat Software Architecture from Jupyter Notebooks](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=K4QN27IKr0g) - A primer on how to structure your Machine Learning projects when using Jupyter Notebooks.\n\n\u003Ca name=\"misc-hubs-and-links\" \u002F>\n\n## Misc. Hubs & Links\n\n- [Hacker News](https:\u002F\u002Fnews.ycombinator.com\u002Fnews) - Maybe how I discovered ML - Interesting trends appear on that site way before they get to be a big deal.\n- [DataTau](http:\u002F\u002Fwww.datatau.com\u002F) - This is a hub similar to Hacker News, but specific to data science.\n- [Naver](http:\u002F\u002Fwww.naver.com\u002F) - This is a Korean search engine - best used with Google Translate, ironically. Surprisingly, sometimes deep learning search results and comprehensible advanced math content shows up more easily there than on Google search.\n- [Arxiv Sanity Preserver](http:\u002F\u002Fwww.arxiv-sanity.com\u002F) - arXiv browser with TF\u002FIDF features.\n- [Awesome Neuraxle](https:\u002F\u002Fgithub.com\u002FNeuraxio\u002FAwesome-Neuraxle) - An awesome list for Neuraxle, a ML Framework for coding clean production-level ML pipelines.\n\n\n\u003Ca name=\"license\" \u002F>\n\n## License\n\n[![CC0](http:\u002F\u002Fmirrors.creativecommons.org\u002Fpresskit\u002Fbuttons\u002F88x31\u002Fsvg\u002Fcc-zero.svg)](https:\u002F\u002Fcreativecommons.org\u002Fpublicdomain\u002Fzero\u002F1.0\u002F)\n\nTo the extent possible under law, [Guillaume Chevalier](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier) has waived all copyright and related or neighboring rights to this work.\n","# [深度学习资源精选](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FAwesome-Deep-Learning-Resources) [![Awesome](https:\u002F\u002Fcdn.rawgit.com\u002Fsindresorhus\u002Fawesome\u002Fd7305f38d29fed78fa85652e3a63e154dd8e8829\u002Fmedia\u002Fbadge.svg)](https:\u002F\u002Fgithub.com\u002Fsindresorhus\u002Fawesome)\n\n这是一份我整理的深度学习（Deep Learning）优质资源清单。它对我学习深度学习非常有帮助，我经常用它来复习相关主题或作为参考资料。\n我（[Guillaume Chevalier](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier)）亲自构建了这份清单，并仔细学习了其中列出的所有内容。\n\n## 目录\n\n- [趋势](#trends)\n- [在线课程](#online-classes)\n- [书籍](#books)\n- [文章与博客](#posts-and-articles)\n- [实践资源](#practical-resources)\n  - [库与实现](#librairies-and-implementations)\n  - [常用数据集](#some-datasets)\n- [其他数学理论](#other-math-theory)\n  - [梯度下降算法与优化](#gradient-descent-algorithms-and-optimization)\n  - [复数与数字信号处理](#complex-numbers-and-digital-signal-processing)\n- [论文](#papers)\n  - [循环神经网络（RNN）](#recurrent-neural-networks)\n  - [卷积神经网络（CNN）](#convolutional-neural-networks)\n  - [注意力机制](#attention-mechanisms)\n  - [其他](#other)\n- [YouTube与视频](#youtube)\n- [其他资源与链接](#misc-hubs-and-links)\n- [许可证](#license)\n\n\u003Ca name=\"trends\" \u002F>\n\n## 趋势\n\n以下是2004年至2017年9月的[Google Trends](https:\u002F\u002Fwww.google.ca\u002Ftrends\u002Fexplore?date=all&q=machine%20learning,deep%20learning,data%20science,computer%20programming)数据趋势图：\n\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fguillaume-chevalier_Awesome-Deep-Learning-Resources_readme_940d3fb12df2.png\" width=\"792\" height=\"424\" \u002F>\n\u003C\u002Fp>\n\n你也可以查看 Andrej Karpathy 的 [新文章](https:\u002F\u002Fmedium.com\u002F@karpathy\u002Fa-peek-at-trends-in-machine-learning-ab8a1085a106) 了解机器学习研究的最新趋势。\n\n我认为深度学习是让计算机更像人类思考的关键技术，具有巨大潜力。许多原本用传统算法难以解决的自动化任务，现在通过深度学习可以轻松实现。\n\n摩尔定律（Moore's Law）描述的计算机硬件指数级进步，如今更多体现在GPU而非CPU上，因为物理限制使得原子晶体管无法无限缩小。我们正在向并行架构发展[[了解更多](https:\u002F\u002Fwww.quora.com\u002FDoes-Moores-law-apply-to-GPUs-Or-only-CPUs)]。深度学习通过GPU利用了这种并行架构。此外，深度学习算法未来可能结合量子计算（Quantum Computing）并应用于脑机接口（machine-brain interface）领域。\n\n智能与认知的本质是极其有趣的探索课题，目前尚未完全被理解。这些技术充满前景。\n\n\u003Ca name=\"online-classes\" \u002F>\n\n## 在线课程\n\n- **[DL&RNN 课程](https:\u002F\u002Fwww.dl-rnn-course.neuraxio.com\u002Fstart?utm_source=github_awesome) - 我创建的深度学习与循环神经网络（RNN）密集课程。**\n- [Coursera 上 Andrew Ng 的机器学习课程](https:\u002F\u002Fwww.coursera.org\u002Flearn\u002Fmachine-learning) - 著名的入门级课程，提供[证书](https:\u002F\u002Fwww.coursera.org\u002Faccount\u002Faccomplishments\u002Fverify\u002FDXPXHYFNGKG3)。讲师：斯坦福大学副教授 Andrew Ng，百度首席科学家，Coursera 联合创始人。\n- [Coursera 上 Andrew Ng 的深度学习专项课程](https:\u002F\u002Fwww.coursera.org\u002Fspecializations\u002Fdeep-learning) - Andrew Ng 新推出的5门深度学习课程系列，使用 Python 而非 Matlab\u002FOctave，完成可获得[专项证书](https:\u002F\u002Fwww.coursera.org\u002Faccount\u002Faccomplishments\u002Fspecialization\u002FU7VNC3ZD9YD8)。\n- [Google 的深度学习课程](https:\u002F\u002Fwww.udacity.com\u002Fcourse\u002Fdeep-learning--ud730) - 中高级课程，涵盖深度学习高级概念，适合掌握基础后进行创意开发。\n- [Georgia Tech 的交易机器学习课程](https:\u002F\u002Fwww.udacity.com\u002Fcourse\u002Fmachine-learning-for-trading--ud501) - 介绍交易领域机器学习基础知识及AI与金融概念。Q-Learning部分尤其精彩。\n- [Hugo Larochelle 的神经网络课程（Sherbrooke大学）](https:\u002F\u002Fwww.youtube.com\u002Fplaylist?list=PL6Xpj9I5qXYEcOhn7TqghAJ6NAPrNmUBH) - Hugo Larochelle 免费提供的在线神经网络课程，我已观看部分内容。\n- [GLO-4030\u002F7030 深度神经网络学习](https:\u002F\u002Fulaval-damas.github.io\u002Fglo4030\u002F) - Laval大学 Philippe Giguère 教授的课程。特别推荐其多头注意力机制（multi-head attention mechanism）的可视化演示，详见[第13周第28页幻灯片](http:\u002F\u002Fwww2.ift.ulaval.ca\u002F~pgiguere\u002Fcours\u002FDeepLearning\u002F09-Attention.pdf)。\n- [深度学习与循环神经网络（DL&RNN）](https:\u002F\u002Fwww.neuraxio.com\u002Fen\u002Ftime-series-solution) - 本领域最密集的加速课程（滚动至页面底部）。\n\n\u003Ca name=\"books\" \u002F>\n\n## 书籍\n\n- [代码整洁之道](https:\u002F\u002Fwww.amazon.ca\u002FClean-Code-Handbook-Software-Craftsmanship\u002Fdp\u002F0132350882) - 回归编程基础！学习如何编写整洁代码以提升职业素养。这本书是我读过的最佳书籍，尽管它与深度学习无关。\n- [程序员的职业素养](https:\u002F\u002Fwww.amazon.ca\u002FClean-Coder-Conduct-Professional-Programmers\u002Fdp\u002F0137081073) - 学习如何作为专业程序员与经理沟通。这对任何编程职业都至关重要。\n- [如何创造思维](https:\u002F\u002Fwww.amazon.com\u002FHow-Create-Mind-Thought-Revealed\u002Fdp\u002FB009VSFXZ4) - 通勤时聆听音频版很合适。这本书激励人们反向工程大脑并思考AI编码。\n- [神经网络与深度学习](http:\u002F\u002Fneuralnetworksanddeeplearning.com\u002Findex.html) - 覆盖神经网络和深度学习核心概念的经典书籍。\n- [深度学习 - MIT出版社书籍](http:\u002F\u002Fwww.deeplearningbook.org\u002F) - 书中包含丰富的数学内容，帮助理解实际深度学习原理。\n- [其他我读过的书籍](https:\u002F\u002Fbooks.google.ca\u002Fbooks?hl=en&as_coll=4&num=100&uid=103409002069648430166&source=gbs_slider_cls_metadata_4_mylibrary_title) - 部分书籍与深度学习关联较弱，但仍与本清单相关。\n\n\u003Ca name=\"posts-and-articles\" \u002F>\n\n## 文章与帖子\n\n- [雷·库兹韦尔的预测](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FPredictions_made_by_Ray_Kurzweil) - 雷·库兹韦尔（Ray Kurzweil）关于中长期未来预测的列表。\n- [循环神经网络的惊人有效性](http:\u002F\u002Fkarpathy.github.io\u002F2015\u002F05\u002F21\u002Frnn-effectiveness\u002F) - 必读文章，由安德烈·卡帕蒂（Andrej Karpathy）撰写——正是这篇文章促使我学习循环神经网络（RNN），它展示了RNN在最基础的自然语言处理（NLP）中的应用效果。\n- [神经网络、流形与拓扑](http:\u002F\u002Fcolah.github.io\u002Fposts\u002F2014-03-NN-Manifolds-Topology\u002F) - 从新视角探讨神经元如何映射信息。\n- [理解长短期记忆网络（LSTM）](http:\u002F\u002Fcolah.github.io\u002Fposts\u002F2015-08-Understanding-LSTMs\u002F) - 解释LSTM单元的内部工作机制，并在结论部分提供了有趣的链接。\n- [注意力机制与增强型循环神经网络](http:\u002F\u002Fdistill.pub\u002F2016\u002Faugmented-rnns\u002F) - 通过视觉动画展示注意力机制的入门介绍。\n- [使用深度学习在Spotify推荐音乐](http:\u002F\u002Fbenanne.github.io\u002F2014\u002F08\u002F05\u002Fspotify-cnns.html) - 在音频聚类方面的绝佳案例——由Spotify实习生撰写的文章。\n- [发布SyntaxNet：世界上最精确的解析器开源](https:\u002F\u002Fresearch.googleblog.com\u002F2016\u002F05\u002Fannouncing-syntaxnet-worlds-most.html) - Parsey McParseface的诞生，一种神经语法树解析器。\n- [改进Inception与TensorFlow中的图像分类](https:\u002F\u002Fresearch.googleblog.com\u002F2016\u002F08\u002Fimproving-inception-and-image.html) - 非常有趣的卷积神经网络（CNN）架构（例如：Inception风格的卷积层在减少参数数量方面具有高效性）。\n- [WaveNet：原始音频的生成模型](https:\u002F\u002Fdeepmind.com\u002Fblog\u002Fwavenet-generative-model-raw-audio\u002F) - 实现逼真语音合成的机器：完美的语音生成。\n- [弗朗索瓦·肖莱的推特账号](https:\u002F\u002Ftwitter.com\u002Ffchollet) - Keras作者——拥有有趣的推文和创新性想法。\n- [Neuralink与大脑的神奇未来](http:\u002F\u002Fwaitbutwhy.com\u002F2017\u002F04\u002Fneuralink.html) - 关于大脑未来和脑机接口的启发性文章。\n- [为开发大型文件的深度学习应用迁移到Git LFS](http:\u002F\u002Fvooban.com\u002Fen\u002Ftips-articles-geek-stuff\u002Fmigrating-to-git-lfs-for-developing-deep-learning-applications-with-large-files\u002F) - 轻松管理私有Git项目中的超大文件。\n- [深度学习的未来](https:\u002F\u002Fblog.keras.io\u002Fthe-future-of-deep-learning.html) - 弗朗索瓦·肖莱对深度学习未来的思考。\n- [通过决策树发现数据背后的结构](http:\u002F\u002Fvooban.com\u002Fen\u002Ftips-articles-geek-stuff\u002Fdiscover-structure-behind-data-with-decision-trees\u002F) - 构建并可视化决策树，推断数据背后的隐藏逻辑。\n- [优化神经网络超参数的Hyperopt教程](http:\u002F\u002Fvooban.com\u002Fen\u002Ftips-articles-geek-stuff\u002Fhyperopt-tutorial-for-optimizing-neural-networks-hyperparameters\u002F) - 学习自动优化超参数空间，而非手动调整。\n- [估计深度神经网络的最优学习率](https:\u002F\u002Fmedium.com\u002F@surmenok\u002Festimating-optimal-learning-rate-for-a-deep-neural-network-ce32f2556ce0) - 一种巧妙的技巧，在完整训练之前估计最优学习率。\n- [注释版Transformer](http:\u002F\u002Fnlp.seas.harvard.edu\u002F2018\u002F04\u002F03\u002Fattention.html) - 有助于理解《Attention Is All You Need》（AIAYN）论文。\n- [图解Transformer](http:\u002F\u002Fjalammar.github.io\u002Fillustrated-transformer\u002F) - 同样有助于理解《Attention Is All You Need》（AIAYN）论文。\n- [通过无监督学习提升语言理解](https:\u002F\u002Fblog.openai.com\u002Flanguage-unsupervised\u002F) - 通过大规模语料的无监督预训练，在多个NLP任务中达到SOTA（最先进）水平。\n- [NLP的ImageNet时刻已到来](https:\u002F\u002Fthegradient.pub\u002Fnlp-imagenet\u002F) - 欢呼NLP的ImageNet时刻。\n- [图解BERT、ELMo等（NLP如何攻克迁移学习）](https:\u002F\u002Fjalammar.github.io\u002Fillustrated-bert\u002F) - 理解NLP的ImageNet时刻所采用的不同方法。\n- [Uncle Bob的面向对象设计原则](http:\u002F\u002Fbutunclebob.com\u002FArticleS.UncleBob.PrinciplesOfOod) - 不仅需要SOLID原则来编写干净代码，更进一步的REP、CCP、CRP、ADP、SDP和SAP原则对于开发需要打包成多个独立模块的大型软件至关重要。\n- [为什么87%的数据科学项目从未投入生产？](https:\u002F\u002Fventurebeat.com\u002F2019\u002F07\u002F19\u002Fwhy-do-87-of-data-science-projects-never-make-it-into-production\u002F) - 数据不可忽视，团队与数据科学家之间的沟通对于正确集成解决方案至关重要。\n- [机器学习项目失败的真正原因](https:\u002F\u002Ftowardsdatascience.com\u002Fwhat-is-the-main-reason-most-ml-projects-fail-515d409a161f) - 聚焦清晰的业务目标，避免算法迭代除非代码非常整洁，并能够判断何时代码“足够好”。\n- [面向机器学习的SOLID原则](https:\u002F\u002Fwww.umaneo.com\u002Fpost\u002Fthe-solid-principles-applied-to-machine-learning) - 将SOLID原则应用于机器学习。\n\n\u003Ca name=\"practical-resources\" \u002F>\n\n## 实用资源\n\n\u003Ca name=\"librairies-and-implementations\" \u002F>\n\n### 库和实现  \n- [Neuraxle（机器学习流水线框架）](https:\u002F\u002Fgithub.com\u002FNeuraxio\u002FNeuraxle) - 用于构建和部署机器学习项目的最佳框架，且兼容大多数框架（例如：Scikit-Learn、TensorFlow、PyTorch、Keras 等）。  \n- [TensorFlow 的 GitHub 仓库](https:\u002F\u002Fgithub.com\u002Ftensorflow\u002Ftensorflow) - 最知名的深度学习框架，兼具高层和底层功能，同时保持灵活性。  \n- [skflow](https:\u002F\u002Fgithub.com\u002Ftensorflow\u002Fskflow) - TensorFlow 的 scikit-learn 风格封装。  \n- [Keras](https:\u002F\u002Fkeras.io\u002F) - 另一个类似 TensorFlow 的深度学习框架，主要面向高层接口。  \n- [carpedm20 的仓库](https:\u002F\u002Fgithub.com\u002Fcarpedm20) - 韩国开发者 Taehoon Kim（carpedm20）实现了许多有趣的神经网络架构。  \n- [carpedm20\u002FNTM-tensorflow](https:\u002F\u002Fgithub.com\u002Fcarpedm20\u002FNTM-tensorflow) - 基于 TensorFlow 的神经图灵机（Neural Turing Machine）实现。  \n- [懒人深度学习](http:\u002F\u002Foduerr.github.io\u002Fblog\u002F2016\u002F04\u002F06\u002FDeep-Learning_for_lazybones) - 使用预训练 CNN（AlexNet 2012）的高层嵌入进行视觉领域的迁移学习教程。  \n- [LSTM 用于人体活动识别（HAR）](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FLSTM-Human-Activity-Recognition) - 我关于使用 LSTM 对时间序列进行分类的教程。  \n- [深度堆叠残差双向 LSTM 用于 HAR](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FHAR-stacked-residual-bidir-LSTMs) - 对前述项目的改进版本。  \n- [序列到序列（seq2seq）循环神经网络（RNN）用于时间序列预测](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002Fseq2seq-signal-prediction) - 我关于预测多通道时间序列的教程。  \n- [Hyperopt 优化 Keras CNN 在 CIFAR-100 上的表现](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FHyperopt-Keras-CNN-CIFAR-100) - 在 CIFAR-100 数据集上自动优化神经网络及其架构。  \n- [我收藏的 ML\u002FDL 仓库](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier?direction=desc&page=1&q=machine+OR+deep+OR+learning+OR+rnn+OR+lstm+OR+cnn&sort=stars&tab=stars&utf8=%E2%9C%93) - GitHub 上有许多优秀的代码示例和项目。  \n- [平滑融合图像块](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FSmoothly-Blend-Image-Patches) - 用于 [U-Net 语义分割](https:\u002F\u002Fvooban.com\u002Fen\u002Ftips-articles-geek-stuff\u002Fsatellite-image-segmentation-workflow-with-u-net\u002F) 的图像块平滑合并工具。  \n- [自治理神经网络（SGNN）：投影层](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FSGNN-Self-Governing-Neural-Networks-Projection-Layer) - 使用此方法，无需训练或加载嵌入即可在深度学习模型中使用文本。  \n- [Neuraxle](https:\u002F\u002Fgithub.com\u002FNeuraxio\u002FNeuraxle) - Neuraxle 是一个用于构建整洁流水线的机器学习（ML）库，提供合适的抽象层次以简化 ML 应用的研究、开发和部署。  \n- [Clean Machine Learning，编码练习](https:\u002F\u002Fgithub.com\u002FNeuraxio\u002FKata-Clean-Machine-Learning-From-Dirty-Code) - 通过实践学习良好的机器学习设计模式。  \n\n\u003Ca name=\"some-datasets\" \u002F>  \n\n### 一些数据集  \n\n这些是我发现的可用于模型开发的有趣资源。  \n\n- [UCI 机器学习仓库](https:\u002F\u002Farchive.ics.uci.edu\u002Fml\u002Fdatasets.html) - 包含大量机器学习数据集。  \n- [康奈尔电影对话语料库](http:\u002F\u002Fwww.cs.cornell.edu\u002F~cristian\u002FCornell_Movie-Dialogs_Corpus.html) - 可用于聊天机器人开发。  \n- [SQuAD 斯坦福问答数据集](https:\u002F\u002Frajpurkar.github.io\u002FSQuAD-explorer\u002F) - 可在线探索的问答数据集，以及表现优异的模型列表。  \n- [LibriSpeech ASR 语料库](http:\u002F\u002Fwww.openslr.org\u002F12\u002F) - 高质量的免费英语语音数据集，性别和说话人分布均衡。  \n- [Awesome 公共数据集](https:\u002F\u002Fgithub.com\u002Fcaesar0301\u002Fawesome-public-datasets) - 优质公共数据集合集。  \n- [SentEval：通用句子表示评估工具包](https:\u002F\u002Farxiv.org\u002Fabs\u002F1803.05449) - 用于在多个数据集（NLP 任务）上基准测试句子表示的 Python 框架。  \n- [ParlAI：对话研究软件平台](https:\u002F\u002Farxiv.org\u002Fabs\u002F1705.06476) - 另一个用于在多个数据集（NLP 任务）上基准测试句子表示的 Python 框架。  \n\n\u003Ca name=\"other-math-theory\" \u002F>  \n\n## 其他数学理论  \n\n\u003Ca name=\"gradient-descent-algorithms-and-optimization\" \u002F>\n\n### 梯度下降算法与优化理论\n\n- [神经网络与深度学习，第2章](http:\u002F\u002Fneuralnetworksanddeeplearning.com\u002Fchap2.html) - 反向传播算法（backpropagation）的工作原理概述  \n- [神经网络与深度学习，第4章](http:\u002F\u002Fneuralnetworksanddeeplearning.com\u002Fchap4.html) - 神经网络可以计算任何函数的可视化证明  \n- [是的，你应该理解反向传播](https:\u002F\u002Fmedium.com\u002F@karpathy\u002Fyes-you-should-understand-backprop-e2f06eab496b#.mr5wq61fb) - 揭示反向传播的注意事项及其在模型训练中的重要性  \n- [人工神经网络：反向传播的数学原理](http:\u002F\u002Fbriandolhansky.com\u002Fblog\u002F2013\u002F9\u002F27\u002Fartificial-neural-networks-backpropagation-part-4) - 数学角度解析反向传播  \n- [深度学习第12讲：循环神经网络与LSTM](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=56TYLaQN4N8) - 详细解释RNN图展开过程及梯度下降算法的潜在问题  \n- [鞍点中的梯度下降算法](http:\u002F\u002Fsebastianruder.com\u002Fcontent\u002Fimages\u002F2016\u002F09\u002Fsaddle_point_evaluation_optimizers.gif) - 可视化不同优化器在鞍点的表现  \n- [平坦景观中的梯度下降算法](https:\u002F\u002Fdevblogs.nvidia.com\u002Fwp-content\u002Fuploads\u002F2015\u002F12\u002FNKsFHJb.gif) - 可视化不同优化器在接近平坦景观的表现  \n- [梯度下降](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=F6GSRDoB-Cg) - 已列出Andrew NG的Coursera课程，但此视频作为梯度下降算法的入门介绍特别相关  \n- [梯度下降：直观理解](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=YovTqTY-PYY) - 前一视频的延伸：添加直观理解  \n- [实践中的梯度下降2：学习率](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=gX6fZHgfrow) - 如何调整神经网络的学习率  \n- [过拟合问题](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=u73PU6Qwl1I) - 过拟合现象及其解决方案的详细解释  \n- [偏差与方差诊断](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=ewogYw5oCAI) - 理解神经网络预测中的偏差与方差问题及其解决方案  \n- [自归一化神经网络](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1706.02515.pdf) - 引入革命性的SELU激活函数  \n- [通过梯度下降学习如何通过梯度下降学习](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1606.04474.pdf) - RNN作为优化器：介绍L2L优化器（元神经网络）\n\n\u003Ca name=\"complex-numbers-and-digital-signal-processing\" \u002F>\n\n### 复数与数字信号处理\n\n信号处理可能与深度学习无直接关联，但研究它有助于在基于信号设计神经网络架构时获得更深入的直觉。\n\n- [窗口函数](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FWindow_function) - 维基百科列出的已知窗口函数 - 注意[Hann-Poisson窗口](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FWindow_function#Hann%E2%80%93Poisson_window)对贪婪爬山算法（如梯度下降）特别重要  \n- [MathBox，图形代数与傅里叶分析工具](https:\u002F\u002Facko.net\u002Ffiles\u002Fgltalks\u002Ftoolsforthought\u002F) - 对傅里叶分析的新视角  \n- [如何折叠朱利亚分形](http:\u002F\u002Facko.net\u002Fblog\u002Fhow-to-fold-a-julia-fractal\u002F) - 涉及复数和波动方程的动画  \n- [用动画走向辉煌，数学与物理运动](http:\u002F\u002Facko.net\u002Fblog\u002Fanimate-your-way-to-glory\u002F) - 物理引擎中的收敛方法及其在交互设计中的应用  \n- [用动画走向辉煌 - 第二部分，数学与物理运动](http:\u002F\u002Facko.net\u002Fblog\u002Fanimate-your-way-to-glory-pt2\u002F) - 使用四元数（处理3D旋转的数学对象）进行旋转和旋转插值的动画  \n- [信号滤波、STFT绘图与拉普拉斯变换](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002Ffiltering-stft-and-laplace-transform) - 信号处理的Python简单演示\n\n\u003Ca name=\"papers\" \u002F>\n\n## 论文\n\n\u003Ca name=\"recurrent-neural-networks\" \u002F>\n\n### 循环神经网络\n\n- [神经网络中的深度学习：概述](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1404.7828v4.pdf) - You_Again对深度学习的总结\u002F概述，主要聚焦RNN  \n- [双向循环神经网络](http:\u002F\u002Fwww.di.ufpe.br\u002F~fnj\u002FRNA\u002Fbibliografia\u002FBRNN.pdf) - 通过时间轴双向扫描提升RNN分类性能  \n- [使用RNN编码器-解码器进行短语表示学习的统计机器翻译](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1406.1078v3.pdf) - 将两个网络组合成seq2seq（序列到序列）编码器-解码器架构。RNN编码器-解码器，1000个隐藏单元。Adadelta优化器  \n- [神经网络的序列到序列学习](http:\u002F\u002Fpapers.nips.cc\u002Fpaper\u002F5346-sequence-to-sequence-learning-with-neural-networks.pdf) - 4层堆叠的LSTM单元（1000个隐藏层），输入句子反转，并在WMT’14英法数据集上使用波束搜索  \n- [探索语言建模的极限](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1602.02410.pdf) - 在字符级CNN基础上使用词级LSTM的递归模型，采用大量GPU资源  \n- [神经机器翻译与序列到序列模型：教程](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1703.01619.pdf) - NMT主题的有趣概述，主要阅读第8部分关于带注意力机制的RNN复习  \n- [通过随机残差学习探索循环神经网络的深度](https:\u002F\u002Fcs224d.stanford.edu\u002Freports\u002FPradhanLongpre.pdf) - 在情感分析案例中，残差连接可能优于堆叠RNN  \n- [像素循环神经网络](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1601.06759.pdf) - 类似Photoshop的“内容感知填充”功能，用于图像缺失区域补全  \n- [循环神经网络的自适应计算时间](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1603.08983v4.pdf) - 允许RNN自主决定计算时长。期待其与神经图灵机的结合效果。关于该主题的交互式可视化可[在此处](http:\u002F\u002Fdistill.pub\u002F2016\u002Faugmented-rnns\u002F)找到  \n\n\u003Ca name=\"convolutional-neural-networks\" \u002F>\n\n### 卷积神经网络（Convolutional Neural Networks）\n\n- [What is the Best Multi-Stage Architecture for Object Recognition?](http:\u002F\u002Fyann.lecun.com\u002Fexdb\u002Fpublis\u002Fpdf\u002Fjarrett-iccv-09.pdf) - 局部对比度归一化（local contrast normalization）的出色应用。\n- [ImageNet Classification with Deep Convolutional Neural Networks](http:\u002F\u002Fwww.cs.toronto.edu\u002F~fritz\u002Fabsps\u002Fimagenet.pdf) - AlexNet，2012 ILSVRC，ReLU激活函数（Rectified Linear Unit）的突破性应用。\n- [Visualizing and Understanding Convolutional Networks](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1311.2901v3.pdf) - 引入\"去卷积网络层\"（deconvnet layer）的概念。\n- [Fast and Accurate Deep Network Learning by Exponential Linear Units](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1511.07289v1.pdf) - ELU激活函数（Exponential Linear Unit）在CIFAR视觉任务中的应用。\n- [Very Deep Convolutional Networks for Large-Scale Image Recognition](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1409.1556v6.pdf) - 通过堆叠多个3x3卷积+ReLU层实现更大感受野的创新设计，包含\"ConvNet配置\"的优秀表格。\n- [Going Deeper with Convolutions](http:\u002F\u002Fwww.cv-foundation.org\u002Fopenaccess\u002Fcontent_cvpr_2015\u002Fpapers\u002FSzegedy_Going_Deeper_With_2015_CVPR_paper.pdf) - GoogLeNet：引入\"Inception\"模块，通过并行不同尺寸的卷积层（带\"same\"填充）在深度维度进行拼接。\n- [Highway Networks](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1505.00387v2.pdf) - Highway网络：残差连接（residual connections）的早期探索。\n- [Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1502.03167v3.pdf) - 批归一化（Batch Normalization, BN）：通过在批次维度进行归一化，再进行可学习的线性缩放和平移。\n- [U-Net: Convolutional Networks for Biomedical Image Segmentation](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1505.04597.pdf) - U型网络（U-Net）：具有跳跃连接的编码器-解码器结构，适用于像素级图像分割。\n- [Deep Residual Learning for Image Recognition](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1512.03385v1.pdf) - 带批归一化的深度残差层（residual layers），即\"如何通过过多层来过拟合任何视觉数据集，并在有足够的数据时使任何视觉模型在识别任务中正常工作\"。\n- [Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1602.07261v2.pdf) - 通过残差连接改进GoogLeNet。\n- [WaveNet: a Generative Model for Raw Audio](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1609.03499v2.pdf) - 基于扩张因果卷积（dilated causal convolutions）的音频生成模型，可生成高质量语音\u002F音乐。\n- [Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1610.07584v2.pdf) - 3D-GANs：用于3D模型生成和家具嵌入空间的算术操作（类似word2vec的词向量算术）。\n- [Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour](https:\u002F\u002Fresearch.fb.com\u002Fpublications\u002FImageNet1kIn1h\u002F) - 极速分布式CNN训练方法。\n- [Densely Connected Convolutional Networks](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1608.06993.pdf) - CVPR 2017最佳论文，DenseNet（密集连接卷积网络）在CIFAR-10\u002F100和SVHN数据集上取得SOTA性能。\n- [The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1611.09326.pdf) - 结合U-Net和DenseNet思想的新型网络，特别适合大规模图像分割任务。\n- [Prototypical Networks for Few-shot Learning](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1703.05175.pdf) - 通过损失函数中的距离度量实现小样本学习（few-shot learning）。\n\n\u003Ca name=\"attention-mechanisms\" \u002F>\n\n### 注意力机制（Attention Mechanisms）\n\n- [Neural Machine Translation by Jointly Learning to Align and Translate](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1409.0473.pdf) - LSTM的注意力机制（attention mechanism）！图示和公式解释对我帮助极大。我在[这里](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QuvRWevJMZ4)做过相关演讲。\n- [Neural Turing Machines](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1410.5401v2.pdf) - 允许神经网络学习算法并处理长期依赖序列的杰出工作。\n- [Show, Attend and Tell: Neural Image Caption Generation with Visual Attention](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1502.03044.pdf) - 在CNN特征图上使用LSTM注意力机制的神奇效果。\n- [Teaching Machines to Read and Comprehend](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1506.03340v3.pdf) - 文本问答领域的创新性突破性研究。\n- [Effective Approaches to Attention-based Neural Machine Translation](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1508.04025.pdf) - 探索不同注意力机制实现方式。\n- [Matching Networks for One Shot Learning](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1606.04080.pdf) - 通过注意力机制和查询机制实现低数据量下的单样本学习。\n- [Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1609.08144.pdf) - 2016年：带注意力机制的堆叠残差LSTM是神经机器翻译（NMT）的最佳选择。\n- [Hybrid computing using a neural network with dynamic external memory](http:\u002F\u002Fwww.nature.com\u002Farticles\u002Fnature20101.epdf?author_access_token=ImTXBI8aWbYxYQ51Plys8NRgN0jAjWel9jnR3ZoTv0MggmpDmwljGswxVdeocYSurJ3hxupzWuRNeGvvXnoO8o4jTJcnAyhGuZzXJ1GEaD-Z7E6X_a9R-xqJ9TfJWBqz) - 基于NTM的可微分记忆改进，现称为可微分神经计算机（DNC）。\n- [Massive Exploration of Neural Machine Translation Architectures](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1703.03906.pdf) - 揭示序列到序列框架下NMT有效性的边界。\n- [Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions](https:\u002F\u002Farxiv.org\u002Fpdf\u002F1712.05884.pdf) - 使用WaveNet作为声码器，结合Tacotron 2 LSTM网络的注意力机制，从文本生成高质量音频。\n- [Attention Is All You Need](https:\u002F\u002Farxiv.org\u002Fabs\u002F1706.03762) (AIAYN) - 引入多头自注意力（multi-head self-attention）和位置编码（positional encoding）的Transformer模型，实现无需RNN\u002FCNN的句子级NLP（必读论文，另见[这篇解析](http:\u002F\u002Fnlp.seas.harvard.edu\u002F2018\u002F04\u002F03\u002Fattention.html)和[可视化教程](http:\u002F\u002Fjalammar.github.io\u002Fillustrated-transformer\u002F)）。\n\n\u003Ca name=\"other\" \u002F>\n\n### 其他\n\n- [ProjectionNet: Learning Efficient On-Device Deep Networks Using Neural Projections](https:\u002F\u002Farxiv.org\u002Fabs\u002F1708.00630) - 在深度神经网络中用词投影（word projections）替代词嵌入（word embeddings），无需预提取词典或存储嵌入矩阵。\n- [Self-Governing Neural Networks for On-Device Short Text Classification](http:\u002F\u002Faclweb.org\u002Fanthology\u002FD18-1105) - 本文是对上方ProjectionNet的后续研究。SGNN（自治理神经网络）基于ProjectionNet进一步扩展，详细说明了优化方法（另见我的[代码复现尝试](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier\u002FSGNN-Self-Governing-Neural-Networks-Projection-Layer)以及观看[演讲录像](https:\u002F\u002Fvimeo.com\u002F305197775)）。\n- [Matching Networks for One Shot Learning](https:\u002F\u002Farxiv.org\u002Fabs\u002F1606.04080) - 通过少量示例（无明确类别）对新样本进行分类，每个分类任务数据量少但存在大量相似任务的数据。该方法似乎比孪生网络（Siamese Networks）更优。简而言之：Matching Networks 可直接优化示例间的余弦相似度（类似于自注意力机制的匹配结果），并将结果直接输入softmax。我认为Matching Networks可能可以用于word2vec的CBOW或Skip-gram模型的负采样softmax训练，而无需进行上下文嵌入查找。\n\n\u003Ca name=\"youtube\" \u002F>\n\n## YouTube 和 视频\n\n- [Attention Mechanisms in Recurrent Neural Networks (RNNs) - IGGG](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=QuvRWevJMZ4) - 关于注意力机制（Paper: Neural Machine Translation by Jointly Learning to Align and Translate）的读书小组讲座。\n- [Tensor Calculus and the Calculus of Moving Surfaces](https:\u002F\u002Fwww.youtube.com\u002Fplaylist?list=PLlXfTHzgMRULkodlIEqfgTS-H1AY_bNtq) - 正确理解张量（Tensor）运算的通用方法，仅观看几段视频即可显著提升概念理解。\n- [Deep Learning & Machine Learning (Advanced topics)](https:\u002F\u002Fwww.youtube.com\u002Fplaylist?list=PLlp-GWNOd6m4C_-9HxuHg2_ZeI2Yzwwqt) - 我整理的深度学习相关视频列表，涵盖各类高级主题。\n- [Signal Processing Playlist](https:\u002F\u002Fwww.youtube.com\u002Fplaylist?list=PLlp-GWNOd6m6gSz0wIcpvl4ixSlS-HEmr) - 我制作的信号处理视频合集，包含DFT\u002FFFT、STFT和拉普拉斯变换等内容（我的软件工程本科课程中几乎没有信号处理内容，除了量子物理课的一小部分）。\n- [Computer Science](https:\u002F\u002Fwww.youtube.com\u002Fplaylist?list=PLlp-GWNOd6m7vLOsW20xAJ81-65C-Ys6k) - 我制作的另一份计算机科学主题视频合集。\n- [Siraj's Channel](https:\u002F\u002Fwww.youtube.com\u002Fchannel\u002FUCWN3xxRkmTPmbKwht9FuE5A\u002Fvideos?view=0&sort=p&flow=grid) - Siraj 提供的快节奏深度学习趣味教学视频。\n- [Two Minute Papers' Channel](https:\u002F\u002Fwww.youtube.com\u002Fuser\u002Fkeeroyz\u002Fvideos?sort=p&view=0&flow=grid) - 对部分研究论文的浅显概述，例如WaveNet或神经风格迁移。\n- [Geoffrey Hinton interview](https:\u002F\u002Fwww.coursera.org\u002Flearn\u002Fneural-networks-deep-learning\u002Flecture\u002Fdcm5r\u002Fgeoffrey-hinton-interview) - Andrew Ng 对Geoffrey Hinton的访谈，后者谈论了他的研究突破并给学生建议。\n- [Growing Neat Software Architecture from Jupyter Notebooks](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=K4QN27IKr0g) - 使用Jupyter Notebooks构建机器学习项目的结构指南。\n\n\u003Ca name=\"misc-hubs-and-links\" \u002F>\n\n## 杂项资源与链接\n\n- [Hacker News](https:\u002F\u002Fnews.ycombinator.com\u002Fnews) - 可能是我发现机器学习的起点 - 该网站上的有趣趋势通常比主流媒体更早出现。\n- [DataTau](http:\u002F\u002Fwww.datatau.com\u002F) - 类似Hacker News的数据科学专题平台。\n- [Naver](http:\u002F\u002Fwww.naver.com\u002F) - 韩国搜索引擎 - 建议配合Google翻译使用。令人惊讶的是，有时深度学习搜索结果和高等数学内容在这里比Google搜索更容易找到。\n- [Arxiv Sanity Preserver](http:\u002F\u002Fwww.arxiv-sanity.com\u002F) - 带TF\u002FIDF功能的arXiv浏览器。\n- [Awesome Neuraxle](https:\u002F\u002Fgithub.com\u002FNeuraxio\u002FAwesome-Neuraxle) - Neuraxle框架的精选资源列表，用于编写生产级机器学习流水线。\n\n\u003Ca name=\"license\" \u002F>\n\n## 授权协议\n\n[![CC0](http:\u002F\u002Fmirrors.creativecommons.org\u002Fpresskit\u002Fbuttons\u002F88x31\u002Fsvg\u002Fcc-zero.svg)](https:\u002F\u002Fcreativecommons.org\u002Fpublicdomain\u002Fzero\u002F1.0\u002F)\n\n在法律允许的最大范围内，[Guillaume Chevalier](https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier) 已放弃本作品的所有版权及相关权利。","# Awesome-Deep-Learning-Resources 快速上手指南\n\n## 环境准备  \n### 系统要求  \n- 操作系统：Windows 10\u002F11、macOS 10.14+ 或 Linux（推荐 Ubuntu 20.04+）  \n- Python 版本：3.7 - 3.11  \n- 可选：NVIDIA GPU（支持 CUDA 11.0+）  \n\n### 前置依赖  \n基础环境：  \n```bash\n# 安装 Python 包管理器（如 pip\u002Fconda）  \n# 推荐使用国内镜像源加速安装  \n# 以 pip 为例：  \npip install --upgrade pip -i https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple  \n```\n\nGPU 加速环境（可选）：  \n```bash\n# 安装 CUDA 工具包（需根据显卡型号选择版本）  \n# 安装 cuDNN（需与 CUDA 版本匹配）  \n```\n\n---\n\n## 安装步骤  \n### 1. 创建虚拟环境（推荐）  \n```bash\n# 使用 conda  \nconda create -n dl_env python=3.9  \nconda activate dl_env  \n\n# 或使用 venv  \npython -m venv dl_env  \nsource dl_env\u002Fbin\u002Factivate  # Linux\u002FmacOS  \ndl_env\\Scripts\\activate     # Windows  \n```\n\n### 2. 安装核心框架  \n```bash\n# PyTorch（推荐使用国内镜像）  \npip install torch torchvision torchaudio -i https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple  \n\n# TensorFlow  \npip install tensorflow -i https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple  \n\n# 其他常用工具  \npip install jupyter matplotlib numpy pandas scikit-learn -i https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple  \n```\n\n---\n\n## 基本使用  \n### 最简 PyTorch 示例  \n```python\nimport torch  \nx = torch.rand(5, 3)  # 创建随机张量  \nprint(x)  \n```\n\n### 最简 TensorFlow 示例  \n```python\nimport tensorflow as tf  \nx = tf.random.normal([5, 3])  # 创建正态分布张量  \nprint(x)  \n```\n\n### 启动 Jupyter Notebook  \n```bash\njupyter notebook  # 直接在浏览器中运行代码示例  \n```\n\n---\n\n> **提示**：若需使用 GPU 加速，请确保已正确安装 CUDA 驱动，并在代码中通过 `torch.cuda.is_available()` 或 `tf.config.list_physical_devices('GPU')` 验证 GPU 支持。","李明是一名刚毕业的软件工程师，计划转行进入人工智能领域，但面对零散的网络资源和复杂的知识体系，他陷入了学习困境。\n\n### 没有 Awesome-Deep-Learning-Resources 时\n- **资源分散难整合**：在多个平台搜索课程时，常遇到重复内容或过时信息，耗费大量时间筛选\n- **学习路径不清晰**：尝试从《深度学习》花书开始，但数学基础薄弱导致前两章就难以推进\n- **实践机会缺失**：虽然完成了Kaggle入门赛，却找不到与课程内容匹配的实战项目\n- **理论实践脱节**：看论文时遇到LSTM结构，但缺乏对应的代码实现参考\n\n### 使用 Awesome-Deep-Learning-Resources 后\n- **一站式资源导航**：通过分类目录直接定位到\"Online Classes\"中的Andrew Ng课程，配合配套的Python实践项目\n- **渐进式学习规划**：按照\"Books\"章节先完成《神经网络与深度学习》的数学基础训练，再进入高级课程\n- **项目驱动学习**：在\"Practical resources\"中找到PyTorch官方教程，边学理论边复现图像分类案例\n- **理论实践闭环**：当研究\"Attention Mechanisms\"论文时，可同步查看HuggingFace的Transformer实现代码\n\n这个工具为深度学习学习者构建了完整的知识图谱，将碎片化资源转化为系统化学习路径，显著提升了从理论掌握到工程落地的效率。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Fguillaume-chevalier_Awesome-Deep-Learning-Resources_33100543.png","guillaume-chevalier","Guillaume Chevalier","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Fguillaume-chevalier_15923461.png","e^(πi) + 1 = 0",null,"Canada","guillaume-chevalier@outlook.com","https:\u002F\u002Fgithub.com\u002Fguillaume-chevalier",1794,294,"2026-04-02T10:45:45","CC0-1.0",1,"","需要 NVIDIA GPU（未说明具体型号），显存需求取决于模型规模（未明确说明）","未说明",{"notes":92,"python":90,"dependencies":93},"未明确说明具体环境配置要求，但部分课程和工具可能需要 CUDA 支持。建议根据实际使用的深度学习框架（如 TensorFlow\u002FPyTorch）查阅其官方文档获取依赖要求。",[],[13],[96,97,98,99,100,101,102],"awesome","awesome-list","deep-learning","machine-learning","tensorflow","lstm","cnn","2026-03-27T02:49:30.150509","2026-04-06T10:26:33.229290",[],[]]