How a neural network mimics the brain to simplify complex data

How a neural network mimics the brain to simplify complex data
Image: Intellect (generated)

In short: Imagine condensing a massive, high-dimensional dataset into a simple, readable two-dimensional map. By using competitive learning rather than error correction, a specialized type of artificial neural network can preserve the essential structure of complex information, making the invisible visible.

Introduced by Finnish professor Teuvo Kohonen in the 1980s, the Self-Organizing Map (SOM) is an unsupervised machine learning technique. Unlike many artificial neural networks that rely on backpropagation and error-correction, the SOM utilizes competitive learning. This approach is inspired by biological models of neural systems from the 1970s and even Alan Turing’s 1950s work on morphogenesis. The goal is to transform a high-dimensional input space into a lower-dimensional, typically two-dimensional, map space while preserving the data's topological structure.

The mechanism relies on a grid of nodes, arranged either hexagonally or rectangularly. During the training phase, the network is fed large numbers of example vectors. For each vector, the system calculates the Euclidean distance to all node weight vectors to identify the 'best matching unit' (BMU). The BMU and its neighbors are then adjusted, or 'pulled,' toward the input vector. This process is governed by a neighborhood function and a learning rate that decreases over time. Initially, the neighborhood is broad, allowing for global organization; as training progresses, the neighborhood shrinks, allowing the nodes to converge on local estimates.

The resulting map functions much like a cortical homunculus—a distorted neurological representation of the human body. In this map, proximal clusters represent observations with similar values, while distal clusters represent different ones. While researchers often use random initialization for node weights, using principal component eigenvectors can accelerate learning. However, the effectiveness of these methods depends on the data's geometry: principal component initialization excels with quasilinear sets, but random initiation may perform better for nonlinear datasets.

Opens the Intellect+ feed with this card first. Swipe up for the next one.

Sources

To cite this card, link to https://intellectplus.co.uk/app/technology-self-organizing-map.

How we source, check and correct cards: Editorial standards.

General information, not professional advice. Spotted an error? Flag the card or see corrections and complaints.

Intellect Plus is a visual educational discovery platform for exploring fascinating facts and ideas. Brain Food, Not Brain Rot. What is Intellect Plus?

Related

Finding something worth knowing…