Research Scientist at Google Brain
Jan 2014 - Dec 2015 (2 years)
I have implemented the first convolutional neural net trained on ImageNet at Google (AlexNet architecture). The training has been distributed on thousands of CPUs as Google had no GPUs back then. The result was so significant that Google moved away from traditional computer vision.
Co-author of "adversarial examples" - imperceptible change to an image that corrupts a neural network.
Author of "Learning to Execute" - demonstrated that neural networks can learn to approximate simple computer programs.