Teaching Machines to See: Convolution, ImageNet, Adversarial Examples and the Apples That Were Bruised on Tuesdays

Deep End Era. Engineering with Sankalp · Oct 28, 2026 · 1h

At the fictional Fenwick Orchards, an apple-grading network with 97% test accuracy starts rejecting sound fruit, but only after the light covers are cleaned. Engineer Arjun Pillai and quality manager Grace Mwangi trace it to a learned shadow shortcut. Along the way: convolution explained as a sliding stencil (with a worked edge-detection example you can follow by ear), the 2012 AlexNet moment, residual connections, YOLO and U-Net, ViT, ConvNeXt and CLIP, the panda that became a gibbon, stop-sign stickers, texture bias, Google Photos 2015, Gender Shades, the Tempe Uber crash, ImageNet label…