Deep learning Papers

In 1943 Warren McCulloch and Walter Pitts described a simplified model of a neuron and showed that networks of such units can carry out logical reasoning, an early link between brain models and computation.

In 1958 Frank Rosenblatt described the perceptron, a model whose connection strengths change with training so that it learns to classify stimuli by trial and error. It became the starting point for trainable neural networks.

A 1969 book by Marvin Minsky and Seymour Papert that proved mathematical limits on what single-layer perceptrons can compute — notably that they cannot compute parity or connectedness — and has long been credited with steering research away from neural networks.

A 1982 paper by John J. Hopfield showing that large networks of simple neuron-like units can collectively store and recall memories. It became known as the Hopfield network, and the work was recognised by the 2024 Nobel Prize in Physics.

In a 1986 Nature paper, Rumelhart, Hinton and Williams showed that backpropagation lets hidden units in a multilayer network learn useful internal features of a task. The method is standard in most neural networks today.

Published in 1992, this paper by Boser, Guyon and Vapnik presents a training algorithm that maximizes the margin between training patterns and the decision boundary, with experiments on optical character recognition.[^1]

Published in 1997 by Sepp Hochreiter and Jürgen Schmidhuber, the LSTM paper introduced a memory cell with gates and a "constant error carousel" that lets recurrent networks learn across long time gaps. It shaped the recurrent models that came before Transformers.

A 2006 paper by Hinton, Osindero and Teh that uses complementary priors and a greedy layer-by-layer algorithm to train deep belief nets, then fine-tunes them; the resulting three-hidden-layer model both modelled and classified handwritten digits well.

A 2009 ICML paper by Rajat Raina, Anand Madhavan and Andrew Y. Ng that used graphics processors to train deep belief networks and sparse coding models many times faster than multicore CPUs.

The 2009 paper introduced ImageNet, a large, cleanly labelled image database organised by the WordNet hierarchy. It later became the basis of an annual benchmark, the ImageNet Large Scale Visual Recognition Challenge.

A 2010 paper by Xavier Glorot and Yoshua Bengio on why deep feedforward networks are hard to train: sigmoid saturation under random initialization and layer Jacobians whose singular values sit far from 1, plus a new initialization scheme that converges substantially faster.

A CVPR 2011 Best Paper that recovers 3D body joint positions from a single depth image without temporal information, by classifying pixels into body parts, and runs at 200 frames per second on consumer hardware.

A 2011 paper by Glorot, Bordes and Bengio reporting that deep networks built from rectifying neurons match or beat tanh networks and reach their best performance without unsupervised pre-training on large labelled supervised tasks.

The 2012 NeurIPS paper that introduced AlexNet, a large convolutional network trained on GPUs. It won the ILSVRC-2012 image classification competition by a wide margin and is widely seen as a turning point for deep learning in computer vision.

Goodfellow, Bengio, and co-authors proposed a generative-model framework with adversarially trained generator and discriminator.

Submitted to arXiv in September 2014 by Dzmitry Bahdanau, Kyunghyun Cho and Yoshua Bengio, this paper let a translation model look back at the whole source sentence while producing each word. This mechanism, called attention, later became a basis for Transformer models.

A 2014 study showing that stacking many small 3×3 convolution filters, up to 16–19 weight layers, improves large-scale image recognition; its models took first place in localisation and second in classification at ImageNet Challenge 2014 and were released publicly.

Ilya Sutskever, Oriol Vinyals and Quoc Le's 2014 paper showed that two LSTMs, one reading the input and one writing the output, could translate sentences and outperform a phrase-based statistical baseline. It set the encoder–decoder pattern behind later translation models.

The 2014 paper that introduced the Inception architecture and its 22-layer GoogLeNet network, which took first place in the ImageNet 2014 classification and detection challenge with a 6.67% top-5 error.

GloVe is a 2014 method that learns word vectors by factorising global word co-occurrence counts. Released together with pre-trained vectors, it became, alongside word2vec, one of the two standard ways of producing static word embeddings.

Adam is a stochastic optimisation algorithm from a 2014 paper by Diederik P. Kingma and Jimmy Ba that sets per-parameter step sizes from running estimates of the gradient's first and second moments, needing little memory and little hyper-parameter tuning.

A 2015 Nature review by Yann LeCun, Yoshua Bengio and Geoffrey Hinton explaining deep learning: models of multiple processing layers that learn layered representations of data, and the results this produced in speech, vision and other fields.

A 2015 Nature paper by Mnih, Kavukcuoglu, Silver and colleagues showed that one deep Q-network agent, given only screen pixels and the game score, learned to play 49 Atari 2600 games at a level comparable to a professional human tester.

A 2016 paper by Barret Zoph and Quoc V. Le that uses a recurrent network trained with reinforcement learning to design neural network architectures automatically, matching human-designed models on CIFAR-10 and beating LSTM on Penn Treebank.

Mask R-CNN is a 2017 framework for object instance segmentation. It detects objects and produces a high-quality mask for each instance by adding a parallel mask branch to Faster R-CNN.

A 2017 paper that introduced the Transformer, a neural network built entirely on attention with no recurrence. It set new machine translation results and became the architecture behind almost every large language model since.

AlphaGo Zero learned Go only by playing against itself, starting from random play, and beat the earlier champion-defeating AlphaGo 100 games to 0 after three days of training. The Nature paper appeared in October 2017.

A 2018 paper by Matthew E. Peters and five co-authors that represents each word with the internal states of a pre-trained deep bidirectional language model, so vectors change with context; adding them to existing models improved results on six NLP tasks.

OpenAI's 2018 paper introducing GPT, a Transformer decoder pre-trained to predict the next token and then fine-tuned for specific tasks. It improved the state of the art on 9 of 12 tasks studied and set the recipe the GPT series followed.

Google's 2018 BERT paper introduced an encoder-only Transformer that reads text in both directions and is pre-trained by hiding words. It set new results on eleven language tasks and made download-and-fine-tune the default workflow.

A 2020 paper showing that a plain Transformer applied directly to sequences of image patches, with no convolution at all, can match state-of-the-art convolutional networks on image classification.

A 2021 paper by Fedus, Zoph and Shazeer that simplifies Mixture-of-Experts routing so that language models can hold up to a trillion parameters while keeping per-example computation roughly constant.

FlashAttention is an exact attention algorithm built around GPU memory traffic: by tiling computation so intermediates stay on chip, it cuts reads and writes to high bandwidth memory, training Transformers faster and enabling longer sequences.

A 2023 paper by Rafailov and colleagues that rewrites the reinforcement-learning-from-human-feedback objective so a language model can be aligned to preferences with a simple classification loss, without a separate reward model or reinforcement learning.

Mamba is a sequence-model architecture introduced by Albert Gu and Tri Dao that replaces attention with selective state space models, giving linear-time scaling, faster inference, and language-model quality competitive with Transformers twice its size.

AlphaFold 3, published in Nature in May 2024, extends structure prediction to biomolecular interactions, reporting greater accuracy than docking tools for protein–ligand binding and than AlphaFold-Multimer v.2.3 for antibody–antigen complexes.

In 1989 LeCun and colleagues at AT&T Bell Laboratories trained a convolutional network with backpropagation to read handwritten zip-code digits from U.S. mail, reaching 5.0% test error. The work showed that the convolutional design could be learned end to end on a real task.

A 1992 technical note by Watkins and Dayan that set out Q-learning, an incremental and computationally light way for an agent to learn optimal action values, and proved it converges to the optimum with probability 1.

Corinna Cortes and Vladimir Vapnik's 1995 paper presents a learning machine for two-group classification that maps inputs nonlinearly into a high-dimensional feature space, constructs a linear decision surface, and extends earlier work from separable to non-separable data.

MNIST is a database of 60,000 training and 10,000 test images of handwritten digits, normalized to 28×28 pixels and described in the 1998 paper by LeCun and colleagues as a shared benchmark for document recognition.

LeCun, Bottou, Bengio and Haffner’s 1998 review argued that learning systems should replace hand-designed features. It showed that convolutional networks such as LeNet-5 outperformed other methods on handwritten digits, and it introduced the MNIST dataset.

Leo Breiman's 2001 paper defines random forests as combinations of tree predictors whose random vectors are sampled independently from the same distribution, analyzes generalization error through tree strength and correlation, and reports error rates competitive with Adaboost and more robust to noise.

Published in 2003 by Yoshua Bengio and colleagues, this paper proposed learning a vector for every word together with a neural network that predicts the next word. It introduced the core idea behind word embeddings and reached lower test perplexity than n-gram models.

CIFAR-10 and CIFAR-100 are labeled image datasets drawn from the 80 million tiny images collection, documented in a 2009 technical report by Alex Krizhevsky, Vinod Nair and Geoffrey Hinton. CIFAR-10 holds 60,000 32×32 colour images in 10 classes.

A 2009 technical report by Alex Krizhevsky that introduces CIFAR-10 and CIFAR-100 and studies multi-layer generative models for learning features from tiny colour images.

A 2010 NIPS paper by Dahl, Ranzato, Mohamed and Hinton applies Deep Belief Nets and a mean-covariance RBM to acoustic modeling, reaching 20.5% phone error on speaker-independent TIMIT phone recognition.

A 2011 workshop paper by Ronan Collobert, Koray Kavukcuoglu and Clement Farabet describing Torch7, a Lua-based numeric computing framework and machine learning library built for flexible experimentation on fast low-level kernels.

A 2011 Journal of Machine Learning Research paper by Collobert, Weston, Bottou, Karlen, Kavukcuoglu and Kuksa, presenting one neural network architecture and learning algorithm for part-of-speech tagging, chunking, named entity recognition and semantic role labeling, trained on mostly unlabeled data with little task-specific engineering.

In 2012, Quoc Le, Andrew Ng and colleagues trained a large unsupervised neural network on 10 million internet images and showed it could detect faces without a single labelled face.

A 2012 article in IEEE Signal Processing Magazine in which four research groups argued that deep neural networks, replacing Gaussian mixture models, gave better acoustic models for speech recognition.

A 2012 NIPS paper by Jeffrey Dean, Andrew Ng and colleagues that introduced DistBelief, a framework for training deep networks across clusters of thousands of machines, together with the Downpour SGD and Sandblaster optimisation procedures.

Submitted to arXiv in January 2013 by Tomas Mikolov and colleagues at Google, this paper introduced CBOW and Skip-gram, two simple architectures that learn word vectors quickly from very large text collections. They are the basis of the word2vec project.

A 2013 paper by Mikolov, Sutskever, Chen, Corrado and Dean presents continuous Skip-gram as an efficient method for learning distributed word and phrase representations, with extensions that improve vector quality and training speed.

R-CNN, posted to arXiv by Girshick, Donahue, Darrell and Malik in November 2013, applied convolutional neural networks to region proposals for object detection and lifted mean average precision on PASCAL VOC 2012 to 53.3%, more than 30% above the previous best[^1].

Zeiler and Fergus ask why large convolutional networks perform so well on ImageNet, introduce a visualization technique for intermediate layers, and use ablation to find an architecture that outperforms Krizhevsky et al. The model also transfers to Caltech-101 and Caltech-256.

A 2013 paper by Mnih, Kavukcuoglu, Silver and colleagues showed a convolutional neural network trained with a Q-learning variant learning to play seven Atari 2600 games directly from raw pixels, beating prior methods on six and human experts on three.

A 2013 paper by Diederik P. Kingma and Max Welling that introduced a stochastic variational inference and learning algorithm for directed latent-variable models with intractable posteriors, making approximate inference practical on large datasets.

A 2014 paper by Srivastava, Hinton, Krizhevsky, Sutskever and Salakhutdinov introducing dropout, a training trick that randomly removes units so large neural networks overfit less.

DeepFace is a 2014 face verification pipeline that rebuilt the alignment and representation stages of the standard face pipeline, reaching 97.35% accuracy on Labeled Faces in the Wild with a nine-layer network of over 120 million parameters.

A 2014 paper from Google researchers that introduced the Inception architecture, whose GoogLeNet model set a new state of the art in ILSVRC 2014 by widening and deepening a network within a fixed computational budget.

A 2014 paper by Graves, Wayne and Danihelka that couples a neural network to an external memory it reads and writes through attention, making the whole machine differentiable and trainable by gradient descent.

A 2014 paper by Vinyals and colleagues that trained a single deep network to turn a photograph into a sentence, reporting a BLEU-1 score of 59 on Pascal against 25 for the previous state of the art.

A 2014 paper presenting an end-to-end deep-learning speech recognition system that removes hand-engineered noise modules and the phoneme dictionary, relying instead on recurrent networks trained across multiple GPUs and synthesized data.

A 2015 paper by Sergey Ioffe and Christian Szegedy that proposed normalising each layer's inputs per training mini-batch, which allowed higher learning rates, looser initialisation and much faster training of deep networks.

A 2015 paper that introduced a Region Proposal Network sharing full-image convolutional features with the detection network, making proposals nearly cost-free and enabling end-to-end training, 5 fps GPU detection, and state-of-the-art accuracy.

The 2015 paper introduced YOLO, which treats object detection as a single regression from image pixels to bounding boxes and class probabilities. One neural network makes predictions in one evaluation, reaching real-time speed.

The 2015 YOLO paper recast object detection as a single neural network's regression over bounding boxes and class probabilities, achieving real-time detection while reducing false positives on empty background.

The 2015 paper introduced residual networks, which add each layer's input back to its output through a shortcut connection. This made networks of over a hundred layers trainable and won first place in several ILSVRC and COCO tasks in 2015.

The 2016 Nature paper describing AlphaGo, which combined deep neural networks with tree search. In October 2015 its distributed version beat the European Go champion Fan Hui 5-0, the first time a program defeated a professional player in full-sized Go without handicap.

A 2016 analysis of residual networks arguing that with identity skip connections and the activation placed after addition, signals travel directly between blocks; a new residual unit built on this eases training, and a 1001-layer network reaches 4.62% error on CIFAR-10.

WaveNet is a deep neural network that generates raw audio waveforms one sample at a time, predicting each sample from all the ones before it. Used for text-to-speech, it was rated by listeners as more natural than the systems it was compared with.

A 2016 paper describing Google's neural machine translation system, GNMT: a deep LSTM encoder-decoder with attention, designed to make end-to-end translation practical by addressing rare words, parallelism and inference cost.

A 2017 paper by Schulman, Wolski, Dhariwal, Radford and Klimov that introduced proximal policy optimization (PPO), a reinforcement learning method that is much simpler to implement and tune and became OpenAI's default RL algorithm.

A 2017 paper by Lin, Goyal, Girshick, He and Dollar that argues dense one-stage detectors are held back by extreme foreground-background imbalance and proposes focal loss, a reshaped cross-entropy that down-weights well-classified examples.

mixup is a training principle from a 2017 paper by Zhang, Cisse, Dauphin and Lopez-Paz that trains networks on convex combinations of pairs of examples and their labels, improving generalisation on image, speech and tabular benchmarks.

A 2017 arXiv paper, later published in Science, reporting that a single self-play reinforcement learning algorithm, AlphaZero, mastered chess, shogi and Go and beat strong specialised programs in each.

ULMFiT is a 2018 transfer-learning method by Jeremy Howard and Sebastian Ruder for fine-tuning language models across NLP tasks; it beat the state of the art on six text-classification tasks and performed well with few labels.

A 2018 paper by Chen, Rubanova, Bettencourt and Duvenaud that replaces a fixed sequence of hidden layers with a neural network describing the derivative of the hidden state, so outputs are computed by an ODE solver — yielding continuous-depth models with constant memory cost.[^1]

XLNet is a generalized autoregressive pretraining method from a 2019 paper. It learns bidirectional contexts by maximizing expected likelihood over all factorization-order permutations and outperforms BERT on 20 tasks under comparable settings.

A replication study of BERT pretraining that measures how much key hyperparameters and training data size matter, finds BERT was significantly undertrained, and shows a properly trained BERT matches or beats later models, reaching state of the art on GLUE, RACE and SQuAD.

MuZero is a DeepMind reinforcement-learning algorithm that learns a model of its environment and plans inside that model, mastering Atari, Go, chess and shogi without being given the rules of those games.

A 2020 paper by Clark, Luong, Le and Manning that pre-trains text encoders by detecting replaced tokens instead of filling in masked ones, matching much larger baselines with far less compute.

A 2020 paper by Jonathan Ho, Ajay Jain and Pieter Abbeel that trained a U-Net to reverse a gradual noising process, making diffusion models a practical rival to GANs and paving the way for DALL-E 2, Imagen and Stable Diffusion.

A 2020 paper by Alexey Dosovitskiy and colleagues showing that a pure Transformer applied directly to sequences of image patches — the Vision Transformer, or ViT — matches state-of-the-art convolutional networks on image classification when pre-trained on large amounts of data.

A 2022 Nature paper by Degrave, Felici and colleagues showing that a deep reinforcement learning controller can learn on its own to command all of a tokamak's magnetic control coils, producing and holding a range of plasma shapes.

A 2022 paper by William Peebles and Saining Xie that replaces the U-Net backbone of diffusion models with a transformer operating on latent patches, showing that image quality improves predictably as compute grows and reaching a state-of-the-art FID of 2.27 on ImageNet 256×256.

GraphCast is a graph neural network trained on four decades of reanalysis data that produces 10-day global weather forecasts at 0.25 degree resolution in under a minute, outperforming the leading operational deterministic systems on most verification targets.

FlashAttention-2 is a 2023 revision of the FlashAttention attention kernel that reorganizes work across GPU thread blocks and warps, roughly doubling speed and reaching up to 73% of the A100's theoretical peak throughput.

A 2023 Nature paper from researchers at Google DeepMind presenting a scaled-up deep-learning approach to predicting stable crystal structures, screening 2.2 million candidates and reporting a large jump in stability-prediction accuracy.

Tri Dao and Albert Gu's paper introduces state space duality, a framework showing that structured state space models and attention variants are closely related, and uses it to derive Mamba-2, whose core layer runs 2 to 8 times faster than Mamba's.

FlashAttention-3 is an attention algorithm for NVIDIA Hopper (H100) GPUs that reaches up to 75% utilization in FP16 and close to 1.2 PFLOPs/s in FP8, using asynchronous execution and low-precision techniques.

SAM 2 is a Meta FAIR foundation model that segments objects in images and videos from simple prompts such as points, boxes or masks, released together with a large video segmentation dataset and training code.

A paper showing that a 32B model fine-tuned on 1,000 carefully chosen questions and run with a simple "budget forcing" trick can beat o1-preview on competition maths by up to 27%, with model, data and code released.

A 2025 paper proposing Absolute Zero, a reinforcement-learning paradigm in which one model writes its own code-reasoning tasks and then solves them, with a code executor as the only judge, reaching state of the art on coding and maths without external training data.

A 2013 paper by Goodfellow, Warde-Farley, Mirza, Courville and Bengio introduces maxout units, whose output is the maximum of a set of inputs, designed to work with dropout, and reports state-of-the-art results on four benchmarks.

A 2013 paper by Graves, Mohamed and Hinton that examines deep recurrent neural networks for speech recognition, arguing for end-to-end sequence training with methods such as connectionist temporal classification.

A 2014 paper by Mehdi Mirza and Simon Osindero that adds conditioning data to both the generator and the discriminator, enabling a GAN to generate samples such as MNIST digits conditioned on class labels.

A 2015 arXiv paper by Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho and co-authors that introduced an image-captioning model with visual attention, able to show where it is looking while writing each word.

Group Normalization is a 2018 paper by Yuxin Wu and Kaiming He that proposes normalizing channels in groups, independent of batch size, to address batch normalization's weakness when batches are small.

A 2019 Google paper that frames every language task as text in, text out, using one encoder-decoder Transformer and one training setup. It also introduced the C4 dataset, and the released T5 models became common research starting points.