U-Net architecture and its skip connections

WHAT THIS TESTS Whether you understand the canonical segmentation architecture and the precise role of its skip connections, including how they differ from ResNet's.

A GOOD ANSWER COVERS U-Net has a symmetric U-shaped design. The contracting path, or encoder, is a typical convolutional stack that repeatedly applies convolutions and downsampling, shrinking spatial size while growing channel depth to capture increasingly abstract context. The expanding path, or decoder, mirrors this, repeatedly upsampling and convolving to restore the original resolution and produce a per-pixel output. The defining feature is the skip connections: at each resolution level, the feature maps from the encoder are concatenated onto the corresponding decoder feature maps before further convolution. Their purpose is to recover the precise spatial localization that downsampling destroys. The deep decoder features carry strong semantics but coarse location, while the encoder skips carry fine-grained, high-resolution detail, so combining them yields sharp, accurately placed boundaries. The skips also ease gradient flow during training.

COMMON WRONG ANSWERS Describing the skips as residual additions like ResNet; U-Net concatenates feature maps along the channel dimension instead. Saying the decoder alone can recover detail; without skips, boundaries blur. Forgetting that encoder and decoder are symmetric. Confusing U-Net with a plain autoencoder, which lacks these resolution-matched skips.

LIKELY FOLLOW-UPS Why concatenate rather than add. Why was U-Net effective with few training images. How does it compare to FCN's skip fusion. What upsampling does the decoder use.

ONE CONCRETE EXAMPLE In biomedical cell segmentation, the encoder downsamples a microscopy image so deep layers know this region is cell membrane, but the exact membrane outline is lost. The matching encoder skip connection, at full resolution, still holds the crisp edge information. Concatenating it into the decoder lets the network place the membrane boundary precisely, which is why U-Net produces tight, accurate masks even when trained on only a small number of annotated images.