Generalized neural-network representation of high-dimensional potential-energy surfaces
Behler and Parrinello's 2007 paper introduced neural-network interatomic potentials, learning quantum-accurate forces orders of magnitude faster than DFT.
Landmark Papers
What the papers actually said - linked to the originals.
644 entries, all primary-sourced
Universal Intelligence: A Definition of Machine Intelligence
Legg and Hutter's 2007 paper distills expert definitions of intelligence into one formal measure across all environments.
Whole Brain Emulation: A Roadmap (Sandberg and Bostrom, 2008)
A 2008 Oxford report maps the technical requirements for scanning and emulating a complete human brain in software.
The Basic AI Drives
Stephen Omohundro's 2008 paper argued almost any goal-driven AI would converge on the same instrumental drives, like self-protection and resource acquisition.
The Missing Memristor Found
The 2008 Nature paper from HP Labs that built the first memristor, a resistor with memory, reviving a 1971 prediction and a key device for analog AI hardware.
Visualizing Data using t-SNE
Van der Maaten and Hinton's 2008 paper introducing t-SNE, the technique that turns high-dimensional data into readable 2D maps.
Collaborative Filtering for Implicit Feedback Datasets
How to build recommenders from clicks, watches, and purchases instead of explicit star ratings.
Isolation Forest
Liu, Ting, and Zhou's Isolation Forest, an anomaly detector that finds outliers by how easily random splits isolate them, with near-linear cost.
The Anatomy of A.L.I.C.E.
Richard Wallace describes A.L.I.C.E. and AIML, the pattern-matching chatbot and markup language that won the Loebner Prize.
Matrix Factorization Techniques for Recommender Systems
The Netflix Prize winners explain how matrix factorization beats nearest-neighbor methods at predicting what users will like.
Quantum Algorithm for Solving Linear Systems of Equations (HHL)
The 2009 Harrow-Hassidim-Lloyd algorithm that solves sparse linear systems in time logarithmic in the matrix size, a key primitive for quantum machine learning.
The Free-Energy Principle: A Unified Brain Theory?
Karl Friston proposed that brains, perception, action, and learning all work to minimize a single quantity called variational free energy.
From RankNet to LambdaRank to LambdaMART: An Overview
Burges's self-contained account of the ranking algorithms that powered commercial web search.
Building Watson: An Overview of the DeepQA Project
IBM's team describes DeepQA, the question-answering architecture behind the Watson system that beat human champions at Jeopardy.
Factorization Machines
Rendle's model that unified matrix factorization and feature-based prediction for sparse recommendation data.
Adaptive Subgradient Methods (AdaGrad)
The 2011 paper introducing AdaGrad, the per-parameter adaptive learning rate that paved the way for RMSprop and Adam.
Product Quantization for Nearest Neighbor Search
Foundational method that compresses high-dimensional vectors into compact codes, enabling fast approximate nearest neighbor search.
Fairness Through Awareness
An early formal paper on algorithmic fairness that defined statistical parity and the idea that similar people should be treated similarly.
Scikit-learn: Machine Learning in Python
The 2011 JMLR paper introducing scikit-learn, the BSD-licensed Python library that became the standard toolkit for classical machine learning.
Multigait soft robot
A robot made entirely of soft rubber, driven by air pressure and inspired by squid and worms, walked and crawled under obstacles with no rigid skeleton.
Random Search for Hyper-Parameter Optimization
Bergstra and Bengio's 2012 JMLR paper showing random search beats grid search for tuning, because few hyperparameters actually matter.
The Digital Twin Paradigm for Future NASA and U.S. Air Force Vehicles
Glaessgen and Stargel's 2012 NASA paper gave the digital twin its canonical definition: a high-fidelity simulation mirroring a physical vehicle's life.
Practical Bayesian Optimization of Machine Learning Algorithms
Snoek, Larochelle, and Adams' 2012 paper that made Bayesian optimization a practical, automatic way to tune ML hyperparameters.
ImageNet Classification with Deep Convolutional Neural Networks (AlexNet)
The 2012 paper by Krizhevsky, Sutskever, and Hinton whose GPU-trained deep network crushed the ImageNet contest and ignited the deep learning revolution.
On the Importance of Initialization and Momentum in Deep Learning
The 2013 paper showing that good initialization plus well-scheduled momentum lets plain SGD train deep and recurrent networks.
Translating Embeddings for Modeling Multi-relational Data (TransE)
TransE models knowledge graph relations as simple vector translations, a compact and scalable way to embed facts.
Efficient Estimation of Word Representations in Vector Space (word2vec)
The 2013 Google paper that introduced word2vec, a fast way to turn words into vectors that capture meaning and even arithmetic-like analogies.
The Materials Project: a materials genome approach
A 2013 commentary launched the Materials Project, an open database of computed properties for known inorganic materials, now used to train AI models.
The Future of Employment (Frey and Osborne)
Frey and Osborne's 2013 study estimated that about 47 percent of US jobs were at high risk of computerisation over the next two decades.
Rich Feature Hierarchies for Object Detection (R-CNN)
The 2013 R-CNN paper brought deep learning to object detection, lifting accuracy on PASCAL VOC by more than 30 percent over prior methods.
Auto-Encoding Variational Bayes (VAE)
Kingma and Welling's 2013 paper introduced the variational autoencoder, a foundational deep generative model trained by gradient descent.
Intriguing Properties of Neural Networks
The 2013 paper that discovered adversarial examples: tiny imperceptible perturbations that make a neural network misclassify an image.
Dropout: A Simple Way to Prevent Neural Networks from Overfitting
The 2014 paper on dropout, the technique of randomly switching off neurons during training to curb overfitting.
The Algorithmic Foundations of Differential Privacy
Dwork and Roth's monograph that consolidated differential privacy into a textbook, defining the Laplace and Gaussian mechanisms and privacy budgets.
DeepWalk: Online Learning of Social Representations
DeepWalk pioneered learning graph node embeddings by treating random walks like sentences in a language model.
Small-Footprint Keyword Spotting Using Deep Neural Networks
Google researchers show a compact deep neural network can detect a wake word on-device, the core technique behind Okay Google and Alexa.
Learning Phrase Representations using RNN Encoder-Decoder (GRU)
The 2014 Cho et al. paper that introduced the gated recurrent unit, a simpler gated RNN cell that rivals the LSTM with fewer parameters.
Generative Adversarial Networks
The 2014 paper by Goodfellow and colleagues that pitted two neural networks against each other to generate realistic images, launching the GAN era.
Experimental evidence of massive-scale emotional contagion through social networks
A 2014 study secretly tuned 689,003 Facebook feeds and found that emotions can spread through what an algorithm shows.
Caffe: Convolutional Architecture for Fast Feature Embedding
The 2014 Berkeley paper introducing Caffe, an early BSD-licensed deep learning framework that powered much of computer vision research.
DeepFace: Closing the Gap to Human-Level Performance in Face Verification
Facebook's 2014 DeepFace paper hit 97.35% on the LFW face benchmark, nearly matching human accuracy with a deep neural network.
RAPPOR: Randomized Aggregatable Privacy-Preserving Ordinal Response
Google's deployed local differential privacy system that collects population statistics from browsers without seeing any individual's true value.
Neural Machine Translation by Jointly Learning to Align and Translate
The 2014 paper by Bahdanau, Cho, and Bengio that introduced attention, letting a translation model look over the whole sentence instead of one fixed summary.
Very Deep Convolutional Networks (VGG)
The 2014 Oxford paper showing that stacking many small 3x3 convolution layers into a deep, uniform network improves image recognition accuracy.
Going Deeper with Convolutions (GoogLeNet / Inception)
The 2014 Google paper introducing the Inception module, which runs several filter sizes in parallel to build a deeper, more efficient network.
GloVe: Global Vectors for Word Representation
The 2014 Stanford paper introducing GloVe, word embeddings learned from global word-word co-occurrence counts instead of local context windows.
Show and Tell: A Neural Image Caption Generator
Show and Tell fed CNN image features into an LSTM to write captions, treating description as translating an image into a sentence.
Deep Speech: Scaling up End-to-End Speech Recognition
Baidu's 2014 Deep Speech replaced hand-built speech pipelines with one deep network trained end to end, robust to noise.