Fittingly, that innovation came later in 2015, when the Deep Residual Network, a submission led by a young Microsoft researcher named Kaiming He, changed the game yet again. Nicknamed “ResNet” for short, it was enormous—a staggering 152 layers—but employed an architectural twist whereby some of those layers could be bypassed during the training phase, allowing different images to direct their influence toward smaller subregions of the network. Although the fully trained system would eventually put all its depth to use, no single training example was obliged to span its entirety. The result
...more
