Deep learning algorithms capable of analysing more complex kinds of data
Geometric deep learning is an emerging field within deep learning. It refers to a variety of techniques which diverge from the traditional deep learning algorithms we’ve seen in use over the past few years. For whilst certain algorithms have been instrumental in tasks such as speech recognition, image classification and image generation – you might be familiar with the likes of Convolutional Neural Networks (CNNs) or Generative Adversarial Networks (GANs) – they can’t cope with certain types of data.
Increasingly, datasets to which we might want to apply machine learning consist of things like 3D objects, networks, and graphs. These more complex structures are known as non-euclidean data, and can’t be processed by traditional deep learning algorithms. Instead, they must be fed into neural networks within the field of geometric deep learning.
Geometric deep learning has important implications as it allows us to extend the techniques of deep learning to a much greater variety of data.
For example, in chemistry, molecules are typically represented digitally in simplified forms, to make them easier to compute. By instead representing molecular data as a network, we can retain important structural information about the molecule when it is analysed by an algorithm. This makes it easier for scientists to classify or make predictions about the molecule, with clear benefits in fields such as the pharmaceutical and chemical industries. Geometric deep learning techniques essentially make it possible to analyse more kinds of data in their native state, rather than representing the data in a simplified form and losing important information.
For more definitions of technology’s key terms, sign up to our weekly newsletter.