Networks That See

Hold up any photograph and your brain instantly knows what is in it — a dog, a street, a birthday cake. That takes under a second. Teaching a computer to do the same thing was considered nearly impossible for decades. Then a class of deep networks called convolutional neural networks cracked the problem, and computer vision changed overnight.

The Problem With Raw Pixels

A digital image is just a grid of pixels. A small photo might be 224 pixels wide by 224 pixels tall, with three color channels (red, green, blue). That is 224 × 224 × 3 = 150,528 numbers per image. If you handed all those numbers directly to a basic neural network, two problems arise immediately. First, the network has no idea that nearby pixels are related — it treats pixel at position (10, 10) and pixel at position (11, 10) as completely unconnected. Second, if the cat in the photo moves one pixel to the right, the network sees a totally different input and might fail to recognize it. A convolutional neural network — CNN for short — solves both problems with a clever trick called a filter.

Picture a small 3×3 magnifying square sliding across the entire image, one step at a time. That square is a filter. As it moves, it multiplies the pixel values under it by a set of learned numbers and adds them up, producing a single output value. Slide the filter across the whole image and you get a new, smaller grid of values called a feature map. One filter might learn to detect horizontal edges. Another learns vertical edges. A third learns diagonal lines. Stack dozens of filters per layer, then stack many layers, and the network builds up from edges to textures to parts to objects — exactly the hierarchy described in Lesson 1. A famous CNN called AlexNet, trained on 1.2 million images in 2012, cut the error rate on a major image-recognition contest nearly in half compared with every previous method. That single result convinced the research world that deep learning was real.

Computer Vision in the Real World

CNNs now power applications you encounter every day. Medical imaging: CNNs scan X-rays and MRI scans to flag potential tumors, sometimes matching or exceeding specialist accuracy. Self-driving cars: cameras feed into CNNs that identify pedestrians, lane markings, traffic signs, and other vehicles in real time. Face unlock: your phone's front camera runs a CNN in milliseconds to confirm your identity. Content moderation: social platforms use CNNs to automatically detect and remove violent or abusive images before a human reviewer even sees them. Each of these uses the same core idea — filters sliding over pixels — scaled up with massive datasets and specialized hardware.