Receptive Fields

Understanding receptive fields unlocks the entire architecture of convolutional networks — the kind of networks that power image recognition, object detection, and even parts of speech and text processing. Without this concept, you cannot grasp why deeper layers 'see' more context, why stride and pooling speed up training, or how to design a network that can recognise large patterns like faces or whole scenes. It directly opens the door to studying Convolutional Layers, Convolution Arithmetic, and Pooling as a Statistic, and it is essential for later topics like CNN visualization and interpretation.

Imagine you are looking at a photograph through a small round hole in a piece of cardboard. When the hole is tiny, you can only see a single pixel — a dot of colour. As you move the hole around, you see different pixels, but never more than one at a time. Now imagine the hole is a bit bigger, say a 3-by-3 square of pixels. You now see a little patch of the image — maybe part of an edge or a corner. The region of the image that you can see through that hole is your 'receptive field' at that moment. In a neural network, each neuron (a little computing unit) is like that hole. The first layer of neurons looks at tiny patches of the input image. The next layer of neurons takes the outputs of the first layer and looks at a slightly larger region of the original image, because each of its inputs already contains info from a small patch. As you go deeper and deeper, each neuron's 'hole' gets bigger and bigger — it 'sees' more and more of the original image. This is how a network goes from recognising simple edges to recognising complex shapes like faces or cats. The receptive field is simply the size of that hole — the region of the input that actually influences a particular neuron's output.

Think of a detective trying to solve a crime by reading a series of reports. The detective doesn't read the whole case file at once; they start with individual witness statements. Each witness statement is like a neuron's input from a small patch of the image. To understand what happened, the detective reads a few statements together, forming a hypothesis about a specific location (that's like a neuron in the next layer combining several patches). Then, to understand a larger part of the story, they combine several of those hypotheses. Each step, they are able to 'see' a bigger picture. Similarly, in a convolutional network, the first layer sees tiny local patterns (edges, colours), the next layer sees slightly larger patterns (corners, textures), and so on, until the final layers see the whole object or scene. This analogy breaks down, though, in one important way: a detective can choose which reports to read and in what order; a convolutional network has no such choice. Each neuron's receptive field is fixed by the size of the filters and the number of layers — it cannot skip around or 'jump' to a distant part of the image unless the concept of dilation (which we'll mention later) is used. It is purely a function of depth and the settings of the convolution layers.

The receptive field of a neuron in a convolutional network is the region of the original input (such as an image) that can influence that neuron's output — it grows in size as you go to deeper layers.

You have no prior knowledge of neural networks, so I'll build from scratch. This concept connects intimately to a few other ideas you'll soon study: Convolutional Layers (where the filters that create receptive fields live), Convolution Arithmetic (which calculates how big the receptive field is based on kernel size, stride, and padding), and Pooling as a Statistic (where pooling layers shrink the image and thus directly affect how fast receptive fields grow). It's also the foundation for understanding why CNN visualisation techniques — like looking at what regions of an image most activated a neuron — make sense. Without receptive fields, you wouldn't know what a neuron is 'looking at'.