The fishing net that taught pictures to keep their clues
In a harbour shed, I watched a net-mender brace a net for rough weather. Each new row didn’t just stitch to the last one. A thin line ran back through every earlier row, so one tug pulled on loads of knots. More links, clearer pull.
Most picture-reading systems are built like a strict relay. Each step passes on only what it just made. Early useful clues can get ignored, so later steps remake the same sort of clue again. When you tug at the end, the first knots barely feel it.
DenseNets flip that habit inside a same-size stretch. Every new step can grab the outputs from all earlier steps, kept side by side like strands laid next to each other. Nothing gets blended into one mushy strand, so old clues stay intact and usable.
That pile could get bulky, so it’s kept tidy. Each step adds only a small bit of fresh thread and reuses the rest. Before doing heavier tying, the strands are squeezed through a narrow guide, so the next knot is quicker and cleaner.
Then the net has to change scale. Between these shared stretches, there’s a tidy handover that shrinks the picture size and often trims the number of strands carried forward. It’s like folding and trimming a panel so it fits the next section without dragging loose ends.
Put next to other well-known deep picture readers, this approach often kept up or did better while carrying fewer stored numbers and doing less work for similar quality. Watching that net, the idea felt spot on: strength can come from reusing old knots, not tying new ones every time.