Topology of a Neural Network

Abstract

Topology of a neural network refers to the way the neurons are connected, and it is an important factor in how the network functions and learns. A common topology in unsupervised learning is a direct map** of inputs to a collection of units that represents categories (e.g., self-organizing maps). A common topology in supervised learning is the fully connected, three-layer, feedforward network (see Backpropagation and Radial Basis Function Networks). In deep learning, however, many different topologies are used. Some models are dozens or even hundreds of layers deep (e.g., residual networks), others include complex recurrent structures (e.g., LSTM networks), and others include attentional structures (e.g., transformers). Much of the performance of the model depends on its topology, and neural architecture search, i.e., the process of determining the correct topology automatically using, for example, reinforcement learning or neuroevolution, has become a machine learning area of its own.