Room 4 · The Awakening · 2015
ResNet: 152-Layer Networks
ResNet: 152-Layer Networks (Kaiming He and team · Microsoft Research, 2015): Shortcuts that make very deep networks possible.
Very deep networks were hard to train. ResNet added shortcut connections that skip layers, and so made 152-layer networks possible. It won ImageNet 2015 with a 3.57% error, less than a quarter of AlexNet's error three years earlier.
By 2015 deep learning had a puzzling problem. Adding more layers to a network should, in theory, never make it worse: the extra layers could simply learn to pass their input through unchanged. In practice, very deep networks often trained worse than shallower ones, even on their training data. The researchers called this the degradation problem.
Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun at Microsoft Research Asia proposed a simple solution in "Deep Residual Learning for Image Recognition." In a residual network, or ResNet, the input to a block of layers is added directly to the block's output through a shortcut connection. So the block does not need to learn the whole transformation, only the difference, the residual, between its input and the desired output. If the best thing to do is nothing, the block can easily learn to add almost nothing.
The shortcuts also give the error signal of backpropagation a direct path back through the network, so it does not fade as it passes through many layers.
With this design the team trained networks with 152 layers, eight times deeper than leading networks of the previous year, and even experimented with more than a thousand. ResNet won the ImageNet classification challenge in 2015 with a top-five error of 3.57 percent, and also won the detection and localization tasks. For comparison, one well-known estimate placed human error on the same classification task at about 5 percent.
The paper received the best paper award at the CVPR 2016 conference and became one of the most cited scientific papers of the century. Its shortcut connections are now everywhere, including inside every Transformer in Room 5.
Ask your AI: find the residual connection in a Transformer block and explain what it does there.
Source: arxiv.org/abs/1512.03385
See it in the museum →← DeepMind Learns to Play Atari · AlphaGo Defeats Lee Sedol →
Cite this: Nazareno, I. (2026). ResNet: 152-Layer Networks (2015). New Mind Museum. Retrieved November 17, 2026, from https://newmindmuseum.com/museum/resnet-152-layer-networks-2015/