How computers are getting better at recognizing images – Knowable Magazine (knowablemagazine.org)

Computer vision has advanced fast: from roughly 60% accuracy on image-labeling tasks 15 years ago to common success rates near 90% today, driven by breakthroughs like AlexNet (a 2012 convolutional neural network that upended ImageNet benchmarks) and, more recently, vision transformers (ViTs) that split images into patches and better integrate information across the scene. But progress isn’t just higher scores — researchers have exposed fundamental weaknesses in traditional nets: tiny, human-imperceptible pixel perturbations or small shifts can make models misclassify objects (the famous “cat → guacamole” adversarial example), a symptom of models learning brittle feature shortcuts rather than object concepts. To close that gap, the field is moving toward architectures that more closely mimic human perception. ViTs improve global reasoning across patches, and object-centric neural networks explicitly represent images as compositions of objects rather than collections of features. Object-centric models generalize better in transfer tests (matching irregular shapes: ~86.4% vs ~65.1% for alternative models) and extend to video reasoning and 3D robotic manipulation—enabling robots to grasp, rotate, open drawers, and even harvest fruit by detecting ripeness and navigating branches. The practical implication: more robust, generalizable vision systems for real-world robotics and multimodal AI—though human-level visual understanding remains an open challenge.