The Model Builds Its Own Representation | Visual AI Learning Guide

Follow the input from left to right. On the far left is the raw data itself: a grid of pixel values, or a stream of audio samples, with no labels attached. It enters the first layer, which transforms it into a new set of numbers. That output feeds the next layer, which transforms it again, and so on through the stack, until the final layer produces the prediction. The important thing to notice is what each intermediate stage represents. The early layers respond to simple local patterns; later layers combine those into larger structures. Nobody wrote those descriptions down in advance. They exist because training adjusted the layers until those intermediate descriptions made the prediction more accurate. So the representation is built by the model, layer by layer, from the raw input upward.

0:00 / 0:00

Auto-play

A deep learning model is handed the raw input and nothing else. For an image, that input is a grid of pixel values; for audio, a stream of amplitude samples; for text, a sequence of tokens. No one tells the model that edges, textures, or shapes are the relevant properties. Instead, the model is built from a stack of layers, and each layer transforms the output of the layer before it. The earliest layers tend to respond to simple, local patterns, later layers combine those into larger structures, and the final layer produces the prediction. This stack is what people mean by an internal representation: a set of intermediate descriptions of the input that the model itself constructs, layer by layer, because those descriptions turned out to reduce the prediction error during training. The crucial consequence is that the human no longer names the features. The human chooses the architecture, the data, and the training procedure, and the model discovers which properties of the raw input matter. This is why a deep model can accept raw pixels or raw audio samples that would be impractical to hand-design, and why the same general recipe can be applied to images, sound, and text without a domain expert specifying the properties in advance.