# 30papers.com--伊利亚精选的30篇机器学习必读论文，以适合初学者的形式呈现 - 来源：Hacker News 热门（buzzing.cc 中文翻译） - 作者：notmcrowley - 发布时间：2026-07-08 02:39 - AIHOT 分数：39 - AIHOT 链接：https://aihot.virxact.com/items/cmrb10euj02t6ihogauwd46y9 - 原文链接：https://30papers.com/ ## AI 摘要 30papers.com 网站汇集了伊利亚·苏茨克维精选的30篇机器学习必读论文，每篇附有贡献者列表和简要说明。内容覆盖从卷积神经网络（CS231n、AlexNet、ResNet）、循环网络（RNN、LSTM）到注意力机制与Transformer等里程碑论文，以适合初学者的形式呈现，适合入门学习。 ## 正文 Contributors Andrej Karpathy Fei-Fei Li +many other researchers CS231n: Convolutional Neural Networks for Visual Recognition The course notes that teach convolutional networks from first principles, from linear classifiers up to deep architectures for images. CS231n: Convolutional Neural Networks for Visual Recognition The course notes that teach convolutional networks from first principles, from linear classifiers up to deep architectures for images. Andrej KarpathyFei-Fei Li Plus many more Contributors Andrej Karpathy The Unreasonable Effectiveness of Recurrent Neural Networks A hands on blog post that trains character level RNNs to generate text and shows, with vivid examples, how much structure they capture. The Unreasonable Effectiveness of Recurrent Neural Networks A hands on blog post that trains character level RNNs to generate text and shows, with vivid examples, how much structure they capture. Andrej Karpathy Contributors Christopher Olah Understanding LSTM Networks The clearest visual explanation of how LSTM gates carry information across long sequences, widely used as a first introduction. Understanding LSTM Networks The clearest visual explanation of how LSTM gates carry information across long sequences, widely used as a first introduction. Christopher Olah Contributors Alex Krizhevsky Ilya Sutskever Geoffrey Hinton ImageNet Classification with Deep Convolutional Neural Networks AlexNet. The convolutional network that won ImageNet by a wide margin and set off the modern deep learning era. ImageNet Classification with Deep Convolutional Neural Networks AlexNet. The convolutional network that won ImageNet by a wide margin and set off the modern deep learning era. Alex KrizhevskyIlya Sutskever Contributors Kaiming He Xiangyu Zhang Shaoqing Ren JS Jian Sun Deep Residual Learning for Image Recognition ResNet. Introduces residual connections that let networks grow to hundreds of layers by learning changes to the input rather than full transformations. Deep Residual Learning for Image Recognition ResNet. Introduces residual connections that let networks grow to hundreds of layers by learning changes to the input rather than full transformations. Kaiming HeXiangyu Zhang Contributors FY Fisher Yu Vladlen Koltun Multi-Scale Context Aggregation by Dilated Convolutions Shows how dilated convolutions expand the receptive field without losing resolution, which sharpened dense prediction tasks like segmentation. Multi-Scale Context Aggregation by Dilated Convolutions Shows how dilated convolutions expand the receptive field without losing resolution, which sharpened dense prediction tasks like segmentation. FY Fisher YuVladlen Koltun Contributors Kaiming He Xiangyu Zhang Shaoqing Ren JS Jian Sun Identity Mappings in Deep Residual Networks A follow up to ResNet that studies why identity shortcuts work so well and proposes a cleaner pre-activation residual block. Identity Mappings in Deep Residual Networks A follow up to ResNet that studies why identity shortcuts work so well and proposes a cleaner pre-activation residual block. Kaiming HeXiangyu Zhang Contributors Wojciech Zaremba Ilya Sutskever Oriol Vinyals Recurrent Neural Network Regularization Shows how to apply dropout to LSTMs correctly, on the non-recurrent connections, so large recurrent models stop overfitting. Recurrent Neural Network Regularization Shows how to apply dropout to LSTMs correctly, on the non-recurrent connections, so large recurrent models stop overfitting. Wojciech ZarembaIlya Sutskever Contributors Dario Amodei RA Rishita Anubhai +many other researchers Deep Speech 2: End-to-End Speech Recognition in English and Mandarin An end to end speech recognition system trained with connectionist temporal classification that worked across two very different languages. Deep Speech 2: End-to-End Speech Recognition in English and Mandarin An end to end speech recognition system trained with connectionist temporal classification that worked across two very different languages. Dario Amodei RA Rishita Anubhai Plus many more Contributors Oriol Vinyals Samy Bengio MK Manjunath Kudlur Order Matters: Sequence to Sequence for Sets Examines how the order of inputs and outputs affects sequence to sequence models, and how to handle data that is really a set. Order Matters: Sequence to Sequence for Sets Examines how the order of inputs and outputs affects sequence to sequence models, and how to handle data that is really a set. Oriol VinyalsSamy Bengio Contributors Dzmitry Bahdanau Kyunghyun Cho Yoshua Bengio Neural Machine Translation by Jointly Learning to Align and Translate Introduces the attention mechanism, letting a translation model look back at the relevant source words instead of a single fixed summary. Neural Machine Translation by Jointly Learning to Align and Translate Introduces the attention mechanism, letting a translation model look back at the relevant source words instead of a single fixed summary. Dzmitry BahdanauKyunghyun Cho Contributors Oriol Vinyals Meire Fortunato Navdeep Jaitly Pointer Networks A sequence model whose outputs point back at positions in the input, which suits problems whose answer is a selection or ordering of the inputs. Pointer Networks A sequence model whose outputs point back at positions in the input, which suits problems whose answer is a selection or ordering of the inputs. Oriol VinyalsMeire Fortunato Contributors Ashish Vaswani Noam Shazeer +many other researchers Attention Is All You Need The Transformer. Replaces recurrence entirely with self attention, the architecture that underpins almost every modern large language model. Attention Is All You Need The Transformer. Replaces recurrence entirely with self attention, the architecture that underpins almost every modern large language model. Ashish VaswaniNoam Shazeer Plus many more Contributors Sasha Rush +many other researchers The Annotated Transformer A line by line, runnable reimplementation of the Transformer that turns the original paper into working, readable code. The Annotated Transformer A line by line, runnable reimplementation of the Transformer that turns the original paper into working, readable code. Sasha Rush Plus many more Contributors Alex Graves GW Greg Wayne ID Ivo Danihelka Neural Turing Machines Couples a neural network to an external memory it can read and write with differentiable attention, learning simple algorithms from examples. Neural Turing Machines Couples a neural network to an external memory it can read and write with differentiable attention, learning simple algorithms from examples. Alex Graves GW Greg Wayne Contributors Adam Santoro DR David Raposo +many other researchers A Simple Neural Network Module for Relational Reasoning Introduces the relation network, a small plug in module that lets a network reason about how pairs of objects relate to each other. A Simple Neural Network Module for Relational Reasoning Introduces the relation network, a small plug in module that lets a network reason about how pairs of objects relate to each other. Adam Santoro DR David Raposo Plus many more Contributors Adam Santoro RF Ryan Faulkner +many other researchers Relational Recurrent Neural Networks Adds a self attention based memory to recurrent networks so that stored memories can interact, improving tasks that need relational reasoning over time. Relational Recurrent Neural Networks Adds a self attention based memory to recurrent networks so that stored memories can interact, improving tasks that need relational reasoning over time. Adam Santoro RF Ryan Faulkner Plus many more Contributors Justin Gilmer SS Samuel S. Schoenholz PF Patrick F. Riley Oriol Vinyals D Dahl Neural Message Passing for Quantum Chemistry Unifies many graph neural networks under a message passing framework and applies it to predicting molecular properties. Neural Message Passing for Quantum Chemistry Unifies many graph neural networks under a message passing framework and applies it to predicting molecular properties. Justin Gilmer SS Samuel S. Schoenholz Contributors Jared Kaplan SM Sam McCandlish +many other researchers Scaling Laws for Neural Language Models Measures how language model loss falls as a smooth power law in model size, data, and compute, the empirical basis for building ever larger models. Scaling Laws for Neural Language Models Measures how language model loss falls as a smooth power law in model size, data, and compute, the empirical basis for building ever larger models. Jared Kaplan SM Sam McCandlish Plus many more Contributors YH Yanping Huang YC Youlong Cheng +many other researchers GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism A pipeline parallelism library that splits a giant model across devices and keeps them busy, making it practical to train very large networks. GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism A pipeline parallelism library that splits a giant model across devices and keeps them busy, making it practical to train very large networks. YH Yanping Huang YC Youlong Cheng Plus many more Contributors Geoffrey Hinton DV Drew van Camp Keeping Neural Networks Simple by Minimizing the Description Length of the Weights An early information-theoretic argument that good networks are ones whose weights can be described with few bits, linking generalization to compression. Keeping Neural Networks Simple by Minimizing the Description Length of the Weights An early information-theoretic argument that good networks are ones whose weights can be described with few bits, linking generalization to compression. Geoffrey Hinton DV Drew van Camp Contributors Peter Grunwald A Tutorial Introduction to the Minimum Description Length Principle A readable introduction to choosing models by how well they compress the data, treating learning as finding the shortest description. A Tutorial Introduction to the Minimum Description Length Principle A readable introduction to choosing models by how well they compress the data, treating learning as finding the shortest description. Peter Grunwald Contributors Scott Aaronson The First Law of Complexodynamics A blog essay asking for a formal law that explains why the complexity of a closed system rises, peaks, and falls, rather than simply tracking entropy. The First Law of Complexodynamics A blog essay asking for a formal law that explains why the complexity of a closed system rises, peaks, and falls, rather than simply tracking entropy. Scott Aaronson Contributors Scott Aaronson Sean M. Carroll LO Lauren Ouellette Quantifying the Rise and Fall of Complexity in Closed Systems: The Coffee Automaton Uses a simple cellular automaton model of coffee mixing with cream to ask why complexity rises and then falls as a system moves toward equilibrium. Quantifying the Rise and Fall of Complexity in Closed Systems: The Coffee Automaton Uses a simple cellular automaton model of coffee mixing with cream to ask why complexity rises and then falls as a system moves toward equilibrium. Scott AaronsonSean M. Carroll Contributors Thomas M. Cover JA Joy A. Thomas Kolmogorov Complexity The textbook treatment of the shortest program that produces a string, the formal backbone behind description length and algorithmic randomness. Kolmogorov Complexity The textbook treatment of the shortest program that produces a string, the formal backbone behind description length and algorithmic randomness. Thomas M. Cover JA Joy A. Thomas Contributors XC Xi Chen Diederik P. Kingma +6 6 more researchers Variational Lossy Autoencoder Combines variational autoencoders with autoregressive decoders, and shows how to control which information the latent code is forced to keep. Variational Lossy Autoencoder Combines variational autoencoders with autoregressive decoders, and shows how to control which information the latent code is forced to keep. XC Xi ChenDiederik P. Kingma Plus 6 more Contributors Shane Legg Machine Super Intelligence A doctoral thesis that proposes a formal, universal measure of machine intelligence and explores its consequences for very capable agents. Machine Super Intelligence A doctoral thesis that proposes a formal, universal measure of machine intelligence and explores its consequences for very capable agents. Shane Legg GitHub