Convolutional Neural Network

A Convolutional Neural Network is a program that learns from examples rather than following fixed rules. It is shown a very large number of images together with the correct answer, for instance thousands of photos labeled “dog” or “cat.” The program adjusts itself repeatedly until it can predict the labels correctly on its own. What makes it special is the way it looks at an image: it never views the whole image at once. Instead, it slides a small search window across the image and examines each section individually. This technique of scanning in small windows is called convolution, which is where the method gets its name.

The breakthrough in image understanding

Before this technique, computers had to be told by hand what to look for to recognize an object. Programmers wrote rules for edges, color areas, or curves, for example. This worked on laboratory images but failed on real photos with shadows, blur, and unusual viewing angles. Convolutional networks made this manual work unnecessary. They find the useful features on their own during training.

The well-known turning point was an image recognition competition in 2012. A network called AlexNet reduced the error rate by around ten percentage points compared to the previous best performance. For research, this was an enormous leap within a single year. After that, almost all teams switched to this architecture. Many see this as the starting point of today’s AI boom.

It is important to distinguish this from today’s chatbots. They work with a different architecture, the Transformer, which handles text and long-range context better. For images, however, convolutional networks are still in use, especially when computing power is limited. They run by the millions in cameras, cars, and medical devices.

From edge strokes to faces

To a computer, an image is just a table of numbers. Each number indicates the brightness of a pixel. The search window, technically called a filter, is a tiny table of numbers, often only three by three fields in size. It moves step by step across the image and computes a value at each position. If the section matches the filter’s pattern, the value is high, otherwise it is low.

A network has many such filters side by side and several layers one after another. The first layer usually finds only simple things: vertical lines, corners, light-dark transitions. The next layer works further with these results and detects circles or grids from them. Further along, eyes, wheels, or letters emerge. The final layer combines everything into an answer, such as “cat, 94 percent confidence.”

Between the layers, the image is regularly shrunk. Out of four neighboring values, the network keeps only the largest. This reduces the amount of data and makes recognition insensitive to small shifts. A face is thus recognized even if it is not exactly in the center of the image. A common misconception is that the network “understands” objects. It merely measures how well the learned patterns match the numbers in the image.

Convolutional networks in everyday life

The facial recognition that unlocks your smartphone is based on this technology. So is the photo app that automatically sorts pictures by people or places. The translation app that reads text from a photo also uses a convolutional network for the first step. In cars, such networks analyze camera images and detect pedestrians or speed limits.

In medicine, convolutional networks examine X-rays and tissue samples for conspicuous spots. They do not replace a doctor but flag areas that deserve closer inspection. In business news, they often appear indirectly, for instance when a manufacturer advertises “AI image processing” for quality control on the factory floor. Behind this is almost always a network of this type.

It is notable where these networks run their computations. They are small enough to run directly on a phone or in a camera, without an internet connection. This saves costs and protects data, because the images never leave the device. This is precisely why convolutional networks, despite newer methods, have not disappeared.