ALCOVE (Attention Learning COVEring map), developed by John Kruschke in 1992, is a connectionist model that integrates exemplar-based category representation with gradient-descent learning of attention weights. It explains not only what categories people learn but how they learn to selectively attend to the most diagnostic stimulus dimensions during the learning process.
Architecture
Output: oₘ = Σⱼ wⱼₘ · aⱼ
Response: P(A|x) = exp(φ·o_A) / Σ exp(φ·oₘ)
Learning: Δwⱼₘ = λ_w · (tₘ − oₘ) · aⱼ (weight update)
Attention: Δαₖ = −λ_α · Σₘ(tₘ − oₘ) · Σⱼ wⱼₘ · aⱼ · c · |xₖ − hⱼₖ| (attention update)
Attention Learning
The critical innovation is that attention weights (αₖ) are learned through backpropagation from the error signal. Dimensions that reduce classification error receive increased attention; dimensions that increase error (or are irrelevant) receive decreased attention. This produces learned selective attention that matches human performance on intradimensional/extradimensional shift tasks, filtering, and condensation problems.
ALCOVE demonstrated that connectionist learning mechanisms could be productively combined with the exemplar representation of the GCM, producing a model that captures both the static patterns of classification behavior and the dynamic learning process.