Artificial neural networks frequently employ competitive learning models to categorize input without relying on labelled data. Competitive learning is a kind of Hebbian learning that enhances the specialization of individual nodes within the network.  

The key distinction between Hebbian learning and competitive learning is the number of activated neurons at any given time. In competitive learning, multiple output neurons may be active simultaneously in a neural network based on Hebbian learning. However, in this scenario, only one output neuron is active at any given time.

Network features

In this method, nodes within the network compete to determine which nodes will respond to a specific subset of the input data. The process commences with an input vector, typically in the form of a dataset. The input is subsequently fed into a network of artificial neurons, wherein each neuron possesses its unique set of weights that function as filters. Every individual neuron calculates a score by performing a dot product operation, multiplying the input vector with the neuron's weight and summing the results. 

Winner-Takes-All approach

Following the computation, the neuron with the highest score, sometimes known as the "winner," is typically updated by adjusting its weights to be closer to the input vector. This approach is commonly known as the "Winner-Takes-All" technique. Over time, neurons undergo specialization as they are refined to better align with input vectors. As a result, clusters of comparable data are formed, allowing for the identification of inherent patterns within the input information. 

Real-world example

To exemplify the application of competitive learning, consider an eCommerce enterprise that aims to divide its consumer base for precise marketing yet lacks any pre-existing labels or segmentation. By inputting customer data, such as purchase history, browsing patterns, demographics, etc., into a competitive learning model, it becomes possible to automatically identify specific clusters, such as high spenders, frequent customers, and discount enthusiasts. It enables the customization of marketing techniques to cater to each cluster's preferences. 

Components

Three fundamental components comprise a competitive learning rule:

• A collection of identical neurons, except synaptic weights that are distributed at random; as a result, these neurons react differentially to a given set of input patterns.
• A restriction on the "strength" of an individual neuron.
• A process by which neurons vie for the privilege of responding to a specified subset of inputs, ensuring that only one output neuron (or one neuron per group) is currently active (i.e. "on"). "Winner-take-all" neurons are those that emerge victorious in a competition.

Consequently, as the individual neurons of the network become adept at specializing in ensembles comprising similar patterns, they transform into "feature detectors" for distinct categories of input patterns. Moreover, recoding sets of correlated inputs to one of a small number of output neurons is a characteristic of competitive networks; this effectively eliminates representation redundancy, a critical component of processing in biological sensory systems.

Conclusion

Competitive learning is particularly suitable for datasets in which the quantity of clusters is predetermined, and the data is distributed uniformly across all clusters. It is effective when straightforward, uniform data partitioning is desired.

Furthermore, hierarchical clustering is an excellent alternative when the optimal number of clusters is unknown or when it is necessary to identify hierarchical relationships within the data. It provides the capability to analyze the data at various degrees of detail.