Convolutional Neural Networks

Hello students, welcome to Kwickprep. Your gallery app can find every photo of a dog, even ones you never tagged. Ordinary neural networks find this very hard. So what is the special trick? Today we learn why images are tough, the layers of a convolutional neural network, how it recognises an object, and where it is used.

First, a short reminder. An artificial neuron takes in some numbers and gives out one number. Neurons are arranged in layers: an input layer, hidden layers, and an output layer. Each connection has a weight, a number the network adjusts while learning. In an ordinary network, each neuron connects to every neuron in the layer before, which is called fully connected.

Now think about feeding an image to an ordinary network. A one hundred by one hundred greyscale image has ten thousand pixel values, so each first-layer neuron needs ten thousand weights. The same image in colour has thirty thousand values, needing thirty thousand weights per neuron. A one megapixel colour photo has thirty lakh values, so a single neuron needs thirty lakh weights.

The number of weights is only the first problem. So many weights make training slow and use huge memory. An ordinary network reads the image as one long list, so it loses which pixels were next to each other. If a cat moves a little to the left, every input changes, and the network may not recognise it. With so many weights, it also tends to memorise training photos, which is called overfitting.

A convolutional neural network, or CNN, solves these problems. Instead of looking at the whole image at once, it looks at small patches, just as you scan a page. It reuses the same small kernel, a tiny grid of weights, across the entire image. That reuse means the network needs far fewer weights. Because the kernel slides everywhere, it finds a pattern like an eye wherever the eye appears.

Here are the layers of a CNN, from top to bottom. The convolution layer slides kernels over the image to make feature maps. The ReLU layer, short for rectified linear unit, changes every negative value to zero. The pooling layer shrinks each feature map while keeping the strongest signals. Flattening turns the small maps into one long list of numbers. The fully connected layer uses that list to choose the final label.

Let us look at each layer closely. A convolution layer has many kernels, and each one learns to detect one feature, like a vertical edge. Each kernel slides over the image, multiplying and adding at every position. The result of one kernel is one feature map, showing where that feature appears. We never type these kernel numbers; the network learns them during training.

Pooling makes a feature map smaller. In max pooling, we split a four by four map into two by two blocks, and keep only the biggest value of each. The top-left block has one, three, four and two, so we keep four. The top-right block keeps two. The bottom-left block also keeps two. Pause and predict the last one: the bottom-right block has five, two, three and four, so we keep five.

Why do we pool? Sixteen values became just four, so later layers have much less work. It keeps the strongest signal in each area, so important features survive. Since only the maximum is kept, a small shift in the object's position changes little. After a few rounds of convolution and pooling, flattening lays out the final small maps as one long list for the fully connected layer.

Now let us see how a CNN recognises an object, step by step. The early layers detect simple things like edges and colour patches, such as the lines of whiskers. The middle layers combine edges into shapes and parts, like a pointed ear or an eye. The deep layers combine parts into whole objects, like a cat's face. The output layer gives a probability for each label, for example cat ninety two percent, dog six percent.

How does the CNN learn the right kernels? We show it thousands of labelled images, like photos marked cat or dog. For each image, the network makes a guess. The guess is compared with the correct label to measure the error. Then all the kernel weights are adjusted a little to reduce that error, and this repeats many times.

CNNs power most computer vision around us. On phones, they unlock with your face and let you search photos by words. In health care, they help doctors screen X-rays and eye scans for disease. In farming, apps identify crop diseases from a photo of a leaf. On roads, they read number plates and help cars keep to their lanes. In security, they spot objects such as unattended bags in CCTV footage.

Board exams often ask you to compare, so let us summarise. An ordinary network looks at the whole image at once, while a CNN looks at small patches. The ordinary network needs very many weights, while a CNN shares a few kernels. The ordinary network loses which pixels are neighbours, while a CNN keeps that layout. And if an object moves, the ordinary network gets confused, while a CNN still finds it.

Let us revise what we learned today. Ordinary networks struggle with images because they need too many weights and lose the layout. A CNN uses convolution, ReLU, pooling, flattening and fully connected layers. Max pooling keeps the biggest value in each block. A CNN recognises objects by building from edges, to parts, to whole objects. CNNs are used in phones, health care, farming and on roads. Try max pooling yourself on a four by four grid of your own numbers.