Computer Vision
How Machines Learned to See — and Why They Still Get Fooled by Stickers
A self-driving car can recognise a pedestrian at 60 metres but be completely bamboozled by a stop sign with a small piece of yellow tape on it.
The Idea
For decades, getting computers to interpret images was considered one of the hardest problems in artificial intelligence. Humans do it effortlessly — you glance at a crowded street and instantly know who is walking, where the kerb is, which dog belongs to which owner. Machines, it turned out, needed something more fundamental than clever code: they needed data, and a lot of it. The breakthrough came with deep learning, specifically a type of neural network called a convolutional neural network (CNN). Rather than being programmed with rules about what a cat looks like, a CNN is shown millions of labelled images and learns to detect increasingly abstract features — edges, then textures, then shapes, then objects — layer by layer. By 2012, a CNN called AlexNet slashed the error rate on a major image recognition benchmark so dramatically that the field effectively reset overnight. But here is what is genuinely strange about how this works: no one fully programs what the machine is looking for. The features it learns to detect are emergent — they arise from optimisation, not design. This is why computer vision systems can be brittle in ways that feel almost absurd. They are pattern-matchers of extraordinary power, but the patterns they latch onto are not always the ones you would expect. A classifier trained to spot wolves consistently got it wrong — until researchers discovered it had actually learned to detect snow in the background, because most training images of wolves featured snowy landscapes.
In the World
In 2017, researchers at MIT and other institutions demonstrated something that has since become a landmark provocation in the field: adversarial attacks. By making tiny, carefully calculated changes to an image — changes invisible to the human eye — they could cause a state-of-the-art classifier to misidentify a school bus as an ostrich with near-total confidence. More alarmingly, the same principle works in the physical world. A team at Carnegie Mellon University printed special patterned glasses and wore them in front of a facial recognition system. The system, instead of seeing a human face, confidently identified the wearer as a specific celebrity — or as no one at all. Researchers at the University of Washington printed stickers that, when placed on a stop sign, caused autonomous vehicle vision systems to read it as a speed limit sign. What these experiments reveal is that computer vision systems are not seeing the world the way we do. They are solving a statistical optimisation problem, and adversarial attacks exploit the gap between statistical pattern-matching and genuine visual understanding. The system has never built a model of what a stop sign *is* — it has learned correlations between certain pixel arrangements and a label. Shift those pixel arrangements in a targeted way, and the label changes entirely, while you and I would notice nothing unusual at all.
Why It Matters
Computer vision is already embedded in consequential decisions — medical imaging diagnostics, border surveillance, parole risk assessment, content moderation at global scale. Understanding that these systems are powerful but brittle, and that their brittleness is structural rather than a bug to be patched, changes how you should think about the confidence placed in them. When a radiologist uses an AI tool to screen for tumours, the question worth asking is not just 'how accurate is it on average?' but 'what kinds of images will break it, and are those the images most likely to appear in the patients I see?' The adversarial examples problem is a symptom of a deeper truth: current computer vision learns to correlate, not to comprehend. It has no concept of a stop sign as a physical object with a social function in a traffic system. That gap — between correlation and comprehension — is one of the defining tensions in AI research right now. Knowing it exists makes you a sharper reader of any headline claiming a machine 'sees' or 'understands' something. It almost certainly recognises. Whether it understands is a genuinely open question.
A Question to Ponder
If a system can perform a task as well as a human — or better — but achieves it through a completely different process, does the difference in process matter, and if so, to whom?
Get a new one of these every morning.
Start learning with Thinkable