Geometric Deep Learning (GDL) is predicated on the principle of leveraging the inherent geometric structures within data, a concept not typically addressed by traditional neural network approaches that assume the data is represented in Euclidean space. GDL algorithms interpret the data's shape and structure, such as social networks, molecular structures, or 3D shapes, and can process this non-Euclidean data efficiently. This method is crucial for tasks where the relational or spatial positioning of the data points significantly impacts the output, allowing for more nuanced pattern recognition and decision-making processes. For example, in the analysis of social networks, GDL can identify communities and predict connections based on the graph structure of the network.
The term "Geometric Deep Learning" began gaining traction around the mid-2010s, although the foundational ideas, such as graph neural networks, have been around since the early 2000s. The increased interest correlates with the rise of data science and AI applications that required analyzing data with complex relationships and structures not suited for traditional deep learning frameworks.
Significant figures in the development of Geometric Deep Learning include Michael Bronstein, Joan Bruna, and Xavier Bresson, among others. These researchers have been instrumental in formalizing the theory and expanding its applications across various disciplines such as computational biology, social network analysis, and computer vision. Their collective work has helped establish a framework for applying neural network methodologies to non-Euclidean domains.