The simulation renders a stack of network layers in 3D, showing data flowing forward through weight transformations while highlighting how skip connections carry a parallel gradient path backward during training.
Choose the network depth and toggle residual connections on or off, then press play to watch forward activations and backward gradients animate through the stack, using the speed slider to slow down the flow.
Depth selector, residual on/off toggle, speed slider, play/pause, rebuild network
The original ResNet paper trained a 1,202-layer network as an experiment — it trained successfully thanks to residual connections, though it slightly underperformed the 110-layer version due to overfitting on a small dataset, showing that trainability and generalization are separate challenges.
The simulation renders a stack of network layers in 3D, showing data flowing forward through weight transformations while highlighting how skip connections carry a parallel gradient path backward during training.
The simulation renders a stack of network layers in 3D, showing data flowing forward through weight transformations while highlighting how skip connections carry a parallel gradient path backward during training.
Choose the network depth and toggle residual connections on or off, then press play to watch forward activations and backward gradients animate through the stack, using the speed slider to slow down the flow.
The original ResNet paper trained a 1,202-layer network as an experiment — it trained successfully thanks to residual connections, though it slightly underperformed the 110-layer version due to overfitting on a small dataset, showing that trainability and generalization are separate challenges.