Computing
Turing to AI
Follow how calculation became computation, then machines, software, networks, and intelligence.
-
1936
Turing defines computation through an abstract machine
Alan Turing
Turing analyzed how a person calculates and defined machines whose symbol-by-symbol operations gave a precise mathematical meaning to an effective procedure.
-
1945
The EDVAC report circulates a stored-program architecture
J. Presper Eckert · John Mauchly · John von Neumann · EDVAC team
The First Draft of a Report on the EDVAC described an architecture with instructions and data stored in the same high-speed memory, drawing on collaborative design work.
-
1948
The Manchester Baby runs a stored program
Frederic Williams · Tom Kilburn · Geoff Tootill
The experimental Manchester machine successfully executed a program held in electronic memory.
-
1948
Shannon establishes a mathematical theory of information
Claude Shannon
Shannon quantified information, channel capacity, and the limits of reliable communication in the presence of noise.
-
1943
McCulloch and Pitts model neural activity with logic
Warren McCulloch · Walter Pitts
They represented simplified neurons as threshold logic units and showed how networks could express logical functions.
-
1950
Turing reframes machine intelligence as an operational question
Alan Turing
Turing proposed the imitation game and discussed learning machines, search, objections, and the prospects of digital computers.
-
1956
The Dartmouth workshop names artificial intelligence as a field
John McCarthy · Marvin Minsky · Claude Shannon · Nathaniel Rochester · Dartmouth workshop participants
A summer research project gathered researchers around the proposal that aspects of learning and intelligence could be precisely described and simulated.
-
1958
Rosenblatt demonstrates a trainable perceptron
Frank Rosenblatt · Cornell Aeronautical Laboratory
The perceptron adjusted connection weights from labeled examples to classify simple visual patterns.
-
1956
Logic Theorist proves symbolic logic theorems
Allen Newell · Herbert Simon · Cliff Shaw
Logic Theorist searched for proofs of propositions from Principia Mathematica using symbolic expressions and heuristics.
-
1986
Backpropagation makes multilayer neural-network learning practical
David Rumelhart · Geoffrey Hinton · Ronald Williams
A widely influential paper showed how multilayer networks could learn useful internal representations through gradient backpropagation.
-
1995
Support vector machines advance statistical learning
Corinna Cortes · Vladimir Vapnik
Support vector machines balanced a wide margin between classes against penalties for training errors. Kernels allowed nonlinear boundaries without explicitly constructing every transformed feature.
-
2006
CUDA opens GPUs to general-purpose parallel programming
NVIDIA CUDA team
CUDA provided a programming model for using graphics processors on workloads beyond rendering.
-
2009
ImageNet provides a large labeled visual benchmark
Fei-Fei Li · Jia Deng · ImageNet team
ImageNet organized millions of labeled images into a large hierarchy and paired data scale with a public recognition challenge.
-
2012
AlexNet transforms large-scale image recognition
Alex Krizhevsky · Ilya Sutskever · Geoffrey Hinton
A deep convolutional network trained on GPUs won the ImageNet challenge by a large margin using rectified units, regularization, and data augmentation.
-
2014
Sequence-to-sequence learning maps variable-length inputs to outputs
Ilya Sutskever · Oriol Vinyals · Quoc Le
An encoder-decoder recurrent network learned end-to-end mappings between sequences for tasks such as machine translation.
-
2014
Neural attention learns where to focus during translation
Dzmitry Bahdanau · Kyunghyun Cho · Yoshua Bengio
The model learned a soft alignment over source words while generating each target word rather than compressing the input into one fixed vector.
-
2017
The Transformer replaces recurrence with attention
Ashish Vaswani · Noam Shazeer · Niki Parmar · Google Brain research team
The Transformer used self-attention and feed-forward layers to model sequence relationships in parallel without recurrent computation.
-
2020
Large language models demonstrate broad few-shot behavior
OpenAI GPT-3 research team
A 175-billion-parameter autoregressive Transformer performed many language tasks from prompts and examples without task-specific gradient updates.
Scroll, swipe, or drag to move through time