Deep learning is an advanced form of machine learning that uses neural networks with many layers. These networks, known as deep neural networks, are able to analyse enormous amounts of data and automatically recognise complex patterns and hierarchical structures within it. Each layer of neurons in the network learns to recognise specific traits – from simple ones such as edges and colours in an image to complex ones such as whole objects or faces. This process makes it possible to achieve high accuracy in tasks such as speech recognition, language translation and autonomous driving.
Put simply, you can picture deep learning as training a very powerful “brain” for a computer. Instead of learning from a few examples, it works through millions of data points, photographs of animals for instance. The first layer learns to recognise only basic lines and shapes. The next builds on that and learns to recognise fur, eyes or ears. The final layers can then assemble all these partial pieces of information and determine with high confidence that the picture shows a cat and not a dog. It is learning step by step, from the smallest detail up to the overall picture.
Deep learning
Technical definition: Deep learning is a subcategory of machine learning based on artificial neural networks with multiple hidden layers (so-called deep neural networks). Its characteristic feature is the hierarchical learning of data representations. While the first layers of the network extract low-level features from the data (edges in an image, say), each subsequent layer combines the outputs of the previous one and learns more abstract and complex representations (textures, parts of objects, whole objects). Algorithms such as backpropagation and optimisation methods such as gradient descent are used to train these models on extensive datasets. Typical architectures include convolutional neural networks (CNNs) for image processing and recurrent neural networks (RNNs) for sequential data.
Etymology and context: The term “deep” refers to the number of layers in a neural network – networks with more than one hidden layer are considered deep. Although ideas about layered models existed earlier, the term “deep learning” was popularised by Geoffrey Hinton and other researchers around 2006. The field’s expansion over the past decade was made possible by the availability of large volumes of data and a significant increase in computing power, especially thanks to graphics processing units (GPUs). Deep learning today dominates many areas of artificial intelligence and finds application in computer vision, natural language processing, autonomous systems and bioinformatics.