He et al. introduced residual networks that enabled training hundreds of layers and won ILSVRC 2015.
Kaiming He and colleagues (Microsoft Research) introduced Residual Networks: shortcuts that ease gradient flow and allow very deep nets without training collapse.
ResNet-50/101 became default vision backbones. Skip connections seeped into countless architectures, including NLP variants.
The paper illustrated a deep-learning pattern: an optimization/architecture trick can unlock a new model scale.
Vision in chatbots enabled reasoning over diagrams, screenshots, and scanned documents in one thread.
NV-Reason-CT (arxiv 2609.27511) pairs a Primus/COLIPRI-style 3D encoder with Qwen3.5-4B and GRPO RL; reports ~0.614 macro-F1 and ~0.871 AUROC on CT-RATE, plus a preliminary ~50% interpretation-time cut claim.
In the late 1980s–90s, LeCun and colleagues applied convolutional networks to zip-code and check recognition.