Computer vision aims to turn images into useful information, such as identifying an object or interpreting a medical scan. Capturing an image is not the same as understanding it: a system must also distinguish what matters in the scene. section 6.2 section 6.2.1
Consider a cat. A child may recognize other cats after seeing only a few examples, including cats of different sizes, breeds, or appearances. Deep learning systems typically need many human-labeled images to learn reliable image labels. Recognizing a familiar pattern still does not establish that a system understands an object’s surroundings or significance. section 6.2 section 6.2.2
The COVID-19 imaging test
During the pandemic, researchers trained models to classify chest images as indicating infection or not. Despite the apparent promise of a rapid diagnosis, a review of 2,212 studies found no model with potential clinical use, citing methodological flaws and underlying biases. A plausible image-recognition task did not become a dependable medical tool. section 6.2.1
Image recognition can also be fragile. In a 2017 experiment, changing just one pixel could fool a trained system on many tested images. When judging a computer-vision application, ask not only whether it identifies patterns, but whether its results hold up in the setting where people will rely on them. section 6.2.2 section 6.2.1
