Category: AI – This Quote

Dzmitry Bahdanau and colleagues challenged the fixed-vector bottleneck and introduced neural attention for translation for today’s AI systems.

Kyunghyun Cho and colleagues proposed an encoder-decoder model and introduced the gated recurrent unit for today’s AI systems.

Jacob Devlin and colleagues trained bidirectional language representations that could adapt to many tasks for today’s AI systems.

Ilya Sutskever and colleagues showed how one neural architecture could map entire sequences to new sequences for today’s AI systems.

Joseph Redmon and colleagues reframed object detection as one fast end-to-end prediction problem for computer vision for today’s AI systems.

Ross Girshick and colleagues joined region proposals with learned convolutional features for object detection for today’s AI systems.

Kaiming He and colleagues used residual connections to make very deep neural networks trainable for today’s AI systems.

Nitish Srivastava and colleagues used dropout to reduce overfitting by training many thinned networks for today’s AI systems.

Karen Simonyan and Andrew Zisserman demonstrated how deeper, uniform convolutional networks improved image recognition for today’s AI systems.

Diederik Kingma and Jimmy Ba introduced Adam as an efficient adaptive optimizer for stochastic training for today’s AI systems.

Sergey Ioffe and Christian Szegedy made normalization part of network architecture to improve deep training for today’s AI systems.

Yoav Freund and Robert Schapire generalized online learning into a decision framework that led to AdaBoost for today’s AI systems.