Machine Learning: Achievements
- Updated
On this page
Deep convolutional networks improve image classification
Demonstrated result · 2012
The ImageNet paper reported substantial classification improvements using a deep convolutional network and GPU training. Its contribution helped establish learned visual representations at scale. A benchmark classification result does not by itself establish robust perception in every lighting condition or safety-critical environment.
ImageNet Classification with Deep Convolutional Neural Networks — Krizhevsky, Sutskever and Hinton / NeurIPS · Source date: 2012 · Checked
Learning from distributed local updates
Demonstrated result · 17 February 2016 preprint
McMahan and colleagues described federated learning through aggregated locally computed updates and evaluated model averaging experimentally. The demonstrated result concerns a particular distributed-training approach. Keeping training records on devices does not by itself prove that updates reveal no private information.
Communication-Efficient Learning of Deep Networks from Decentralized Data — McMahan and colleagues / arXiv · Source date: 2016-02-17 · Checked
Attention provides a new sequence architecture
Demonstrated result · 12 June 2017 preprint
Attention Is All You Need introduced the Transformer and evaluated it on translation tasks. Removing recurrence made a different parallel training design possible. Later applications should be evaluated on their own evidence rather than attributing every modern language capability to the original experiments.
Attention Is All You Need — Vaswani and colleagues / arXiv · Source date: 2017-06-12 · Checked