Development of Skip Connection in Deep Neural Networks for Computer Vision and Medical Image Analysis: A Survey — mentioned by D-Machine

Thu, 18 Sep 2025 21:39:08 GMT

I'm not sure your claim is correct, nor particularly precise. Skip connections do propagate feature maps from earlier to later/deeper in the network, and one can model this mathematically as using an identity function to do this, but I wouldn't say that this is "about preserving identity". Also, the reason the identity is useful here is because it prevents vanishing gradients (due to the identity function having an identity tensor as the gradient). Arguably, preventing vanishing gradients is exactly about numerical stability, so I would in fact argue that skip connections are there to make…

Paper and discussion history. A mention is not an endorsement.

Development of Skip Connection in Deep Neural Networks for Computer Vision and Medical Image Analysis: A Survey · HackerPapers discussions

Development of Skip Connection in Deep Neural Networks for Computer Vision and Medical Image Analysis: A Survey — mentioned by D-Machine