Contents

Timeline [16]

A lineage of ideas.

From the rule for updating belief to the architecture behind modern language models — the through-line the atlas follows.

  1. 1763

    Bayes' theorem

    Thomas Bayes (posthumously, communicated by Richard Price); generalized by Pierre-Simon Laplace

    Bayes' essay, published in 1763 after his death, gave a rule for updating the probability of a hypothesis given evidence; Laplace independently restated and developed it into the general form P(HD)=P(DH)P(H)/P(D)P(H\mid D)=P(D\mid H)P(H)/P(D) in the following decades.

  2. 1807

    Fourier series

    Joseph Fourier

    Fourier claimed (in an 1807 memoir, expanded in his 1822 treatise on heat) that any periodic function can be represented as a sum of sines and cosines, founding harmonic analysis and the spectral methods central to signal processing.

  3. 1812

    Least squares and the Gaussian distribution

    Carl Friedrich Gauss and Adrien-Marie Legendre; Laplace

    Legendre published least squares (1805) and Gauss tied it to the normal error distribution and gave a probabilistic justification (1809); Laplace's 1812 treatise unified these with the central limit theorem, establishing the statistical foundations of regression.

  4. 1913

    Markov chains

    Andrey Markov

    Markov introduced chains of dependent random variables in which the next state depends only on the current state, providing the formalism later used in HMMs, MCMC sampling, and reinforcement learning.

  5. 1943

    McCulloch-Pitts artificial neuron

    Warren McCulloch and Walter Pitts

    They proposed a mathematical model of a neuron as a threshold logic unit, showing networks of such units can compute logical functions and laying the conceptual basis for neural networks.

  6. 1958

    The perceptron

    Frank Rosenblatt

    Rosenblatt introduced the perceptron, a trainable linear classifier with a learning rule that adjusts weights from labeled examples, the first practical learning algorithm for a neural model; its linear-separability limits were later analyzed by Minsky and Papert (1969).

  7. 1963

    The Lorenz attractor

    Edward Lorenz

    While studying a simplified model of atmospheric convection, Lorenz found that deterministic nonlinear systems can exhibit sensitive dependence on initial conditions, founding chaos theory and clarifying limits of long-term prediction.

  8. 1970

    Reverse-mode automatic differentiation

    Seppo Linnainmaa

    Linnainmaa described the efficient reverse accumulation of derivatives through a computation graph, the algorithm that underlies backpropagation and all modern gradient-based deep learning frameworks.

  9. 1986

    Backpropagation popularized

    David Rumelhart, Geoffrey Hinton, and Ronald Williams

    Their Nature paper demonstrated that backpropagation could train multilayer networks to learn useful internal representations, reviving neural network research after the perceptron's limitations.

  10. 1989

    Convolutional networks (LeNet)

    Yann LeCun and collaborators

    LeCun applied backpropagation to convolutional networks for handwritten digit recognition (1989), culminating in LeNet-5 (1998), which used weight sharing and local receptive fields to exploit image structure.

  11. 1997

    Long short-term memory (LSTM)

    Sepp Hochreiter and Jurgen Schmidhuber

    LSTM introduced a gated memory cell that mitigates the vanishing-gradient problem, enabling recurrent networks to learn long-range dependencies in sequences.

  12. 2012

    AlexNet

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton

    A deep convolutional network trained on GPUs won the ImageNet ILSVRC-2012 competition by a large margin, igniting the modern deep learning era in computer vision.

  13. 2013

    word2vec

    Tomas Mikolov and colleagues at Google

    word2vec learned dense vector embeddings of words via the skip-gram and CBOW objectives, capturing semantic relationships as linear structure (e.g. king - man + woman ~ queen) and popularizing distributed word representations.

  14. 2014

    Attention for neural machine translation

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio

    They added a soft attention mechanism to encoder-decoder RNNs, letting the decoder dynamically weight source positions and removing the fixed-length bottleneck, the conceptual precursor to the Transformer.

  15. 2014

    Generative adversarial networks (GANs)

    Ian Goodfellow and colleagues

    GANs frame generative modeling as a minimax game between a generator and a discriminator, producing high-fidelity samples without an explicit likelihood.

  16. 2017

    The Transformer

    Ashish Vaswani and colleagues at Google

    The paper 'Attention Is All You Need' replaced recurrence and convolution with pure self-attention, enabling massive parallel training and becoming the backbone of subsequent large language models such as BERT (2018) and the GPT series (2018-).