conceptor

A conceptor is a data-driven linear operator used in recurrent neural networks to represent, control, and manipulate the dynamical patterns a network can generate. Developed within the framework of reservoir computing by Herbert Jaeger, a conceptor is typically a matrix that encodes the correlation structure of reservoir states when the network is driven by a particular signal or pattern. The idea is to capture the subspace of activity associated with that pattern so that the network can be guided to reproduce it or be filtered to suppress other dynamics.

Conceptor construction relies on analyzing the hidden states of a recurrent reservoir during exposure to a pattern. The resulting conceptor matrix acts as a soft constraint on the network’s dynamics: applying the conceptor shapes the state trajectory to stay within the subspace characteristic of the chosen pattern, enabling robust retrieval even in the presence of noise or interference. An aperture parameter controls the balance between specificity and generalization: a larger aperture broadens the subspace, allowing pattern variants to be recalled but potentially increasing false activations.

Boolean operations provide compositional capabilities. OR combines two conceptors to create a union of patterns, AND yields an intersection, and NOT forms the complement, enabling operations such as pattern exclusion or refinement. This enables tasks like memory switching, pattern morphing between concepts, and multi-pattern retrieval.

Applications have included synthetic pattern recall, time-series prediction, and pattern recognition in noisy environments, often within echo state or reservoir computing setups. The conceptor framework has influenced related work on subspace representations and controllable neural dynamics. See also: reservoir computing, echo state networks, pattern recognition.