Hand-calculating U-Net in 17 steps: convs, pooling and skip connections on a 3x16 image
ProfTomYeh · x · 2026-09-18
Prof. Tom Yeh published a new "by hand" tutorial walking through an entire U-Net forward pass in 17 steps on a tiny 3×16 RGB image.
- Encoder path: two 2×3 convolutions each followed by 1×2 max pooling, shrinking 16 columns to 8 then 4, down to a 2×4 bottleneck
- Decoder path: transposed convolutions spread and fill the features, with skip connections copying encoder feature maps across to recover detail lost by pooling
- A final convolution plus ReLU produces R', G', B' outputs at the original size, with full matrices shown for verification
He frames U-Net — born in medical imaging rather than a top conference — as the first image-to-image model that really worked, and now the backbone of modern diffusion-based image generation.
More from Research
- Custom BB-SLAM hits 0.02m accuracy with pure visual odometry, no LiDAR or IMU — broodsugar · 2026-09-18
- Genomics researcher: nobody has 'solved' variant prioritization, but old-score-only views are outdated — anshulkundaje · 2026-09-18
- Aggregate variant scores like CADD still useful when paired with context-specific effects — anshulkundaje · 2026-09-18
- Learning phenotypic scores to map molecular effects into disease-relevant prioritization — anshulkundaje · 2026-09-18
- Collapsing context-specific variant scores into one universal score loses key information — anshulkundaje · 2026-09-18
- Supervised sequence-to-omics models excel at context-specific regulatory predictions — anshulkundaje · 2026-09-18