Deep Learning Papers

ImageNet Classification with Deep Convolutional Neural Networks

The deep-learning breakthrough on ImageNet: a large GPU-trained convolutional network cut top-5 error nearly in half and touched off the modern era of deep learning in vision.

Generative Adversarial Networks

Framed generative modeling as a two-player game between a generator and a discriminator, each improving against the other. GANs drove a decade of image-synthesis research before diffusion models took the lead.

Sequence to Sequence Learning with Neural Networks

Framed translation as sequence-to-sequence learning: an LSTM encoder compresses the source into a vector that an LSTM decoder unrolls into the target. The encoder–decoder skeleton everything that follows hangs off of.

Neural Machine Translation by Jointly Learning to Align and Translate

Introduced attention to neural machine translation: instead of squeezing the source into one vector, the decoder learns to align to and weight over all source positions — the idea the Transformer would later build its whole architecture on.

Adam: A Method for Stochastic Optimization

Adam combined momentum with per-parameter adaptive learning rates into an optimizer that just works across a huge range of problems. It is the default optimizer behind most of the papers in this dossier.

Diederik P. Kingma

Geoffrey Hinton

Ilya Sutskever

Yoshua Bengio

Dzmitry Bahdanau

Convolutional Neural Networks

Layered convolutional filters that learn spatial feature hierarchies — the backbone of modern computer vision.

Sequence-to-Sequence Learning

Mapping one variable-length sequence to another with an encoder–decoder, the framing behind neural translation and beyond.

Pretraining & Transfer Learning

Training a large model on a broad corpus once, then adapting it to many downstream tasks — the paradigm behind BERT and GPT.

Generative Models

Learning to generate data — images, text, audio — by modeling its distribution, often through an adversarial game.

Attention Mechanism

Letting a model weight over all positions of its input, so it can focus on the relevant parts rather than a fixed-size summary.