Modular Feedback Loops and Dynamic Residual Networks
Deep learning textbooks put a plain stack of layers and a residual network in the same box, feedforward, and explain the skip connection as a fix for vanishing gradients. Tracing the plus sign in "next state equals this state plus a correction" from the single premise shows a different cut: what the skip carries is a retained reference, and with it one long, global feedback loop comes apart into many local corrections.
The source is He, Zhang, Ren and Sun, "Deep Residual Learning for Image Recognition" (2015). Kept from it: the degradation problem (a deeper plain network with higher error on…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.