What Are Perception Evolution Networks?

Perception evolution networks are biologically inspired self-organizing neural networks designed for unsupervised, online incremental learning in which the set of sensory inputs itself can grow over time. The model, usually abbreviated PEN, was introduced by researchers at Nanjing University as an extension of the self-organizing incremental neural network family. Its distinguishing property is that new input channels, treated as newly emerged sensory receptors, can be attached to the network during learning, and the knowledge already acquired in the lower-dimensional input space is carried forward and extended into the enlarged space rather than discarded.

The problem PEN addresses sits between clustering and lifelong learning. Conventional unsupervised networks assume a fixed feature vector length and a fixed number of prototypes, so adding a sensor means retraining from scratch. A biological perceptual system does not work that way: an organism that gains a new sense integrates it with what it already knows about the world. PEN formalizes that behavior as a network operation, which is why the original work was framed around adapting to the emergence of a new sensory receptor.

Two-Layer Architecture

A perception evolution network is organized as a perception layer and a knowledge layer. The perception layer receives raw signals from the environment through a set of input channels and passes them upward. The knowledge layer stores learned structure as a graph of prototype nodes connected by edges that record topological neighborhood relations among the observed data. Each input channel in the perception layer projects to the knowledge layer through its own data transmission pathway, and it is this explicit channel-to-pathway mapping that makes structural growth tractable: when a channel is added, the network creates a corresponding pathway instead of rebuilding the representation.

Incremental Prototype Learning

Learning proceeds one sample at a time. For each input the network finds its nearest and second-nearest prototype nodes, decides whether the sample is close enough to be absorbed into the existing representation, and otherwise inserts a new node. Nodes accumulate local statistics that let the network compute a similarity threshold from the data itself, so neither the number of prototypes nor a global distance threshold has to be specified in advance. Edges between winning nodes are created and aged, and stale edges and low-density nodes are pruned, which keeps the topology tracking the current data distribution and gives the method resistance to noise. This class of behavior descends from the self-organizing incremental neural network and from competitive Hebbian learning with growing neural gas, and it is what allows training to continue indefinitely on a stream without a separate consolidation phase.

Cognition Deepening and Receptor Growth

The mechanism that separates PEN from ordinary incremental clustering is its cognition deepening model, described in the IEEE Transactions on Neural Networks and Learning Systems paper on the method. When a new receptor appears, the existing prototypes occupy a subspace of the new, higher-dimensional feature space and carry no information about the added dimension. The network therefore treats the prior prototypes as a scaffold and uses incoming samples to resolve where along the new axis each piece of prior knowledge belongs, splitting a prototype when the new dimension reveals that what looked like one cluster is actually several. The reported experiments, indexed with the published abstract in PubMed, evaluate this on artificial distributions and on multi-attribute pattern data, measuring how quickly the enriched representation recovers accuracy relative to retraining from scratch.

Applications

Perception evolution networks have applications in a range of fields, including:

• Online clustering and topology learning for streaming data
• Robot perception where sensors are added or replaced during operation
• Lifelong and continual machine learning research
• Multimodal sensor fusion in wireless sensor networks
• Unsupervised pattern recognition and novelty detection
• Computational models of perceptual development