# Andreas Veit ResNet lesion study (Cornell, 2016). Topic: [[Topics/LLMs/LLMs|LLMs]] · [[Topics/LLMs/Sources|Sources]] - Lead author, with Michael Wilber and Serge Belongie (Cornell Tech), of "Residual Networks Behave Like Ensembles of Relatively Shallow Networks" (NeurIPS 2016) ([arXiv](https://arxiv.org/abs/1605.06431)). - Deleting single blocks from a 110-layer CIFAR ResNet (or a 200-layer ImageNet one) barely hurts it, except for downsampling blocks; doing the same to VGG drops it to chance. Welch and the post say "56-layer"; the paper's model had 110 layers.