SUSTAIN (Supervised and Unsupervised STratified Adaptive Incremental Network), developed by Love, Medin, and Gureckis (2004), is a clustering model of category learning that adaptively creates new representational units (clusters) in response to surprising or misclassified stimuli. It bridges the gap between supervised learning (learning with feedback) and unsupervised learning (learning the structure of the input).
Cluster Recruitment
2. Compute similarity to each existing cluster
3. Activate the most similar cluster
4. Generate prediction from the activated cluster
5. If prediction is wrong (or stimulus is surprising):
→ Recruit a new cluster centered on x
6. If correct: update the winning cluster toward x
Properties
SUSTAIN starts with a single cluster and adds more as needed to accommodate the category structure. Simple categories (linearly separable) require fewer clusters; complex categories (XOR, family resemblance with exceptions) require more. This adaptive complexity provides a natural account of the difficulty ordering of category structures, the advantage of rule-describable categories over information-integration categories, and the effects of interleaved vs. blocked presentation on category learning.
SUSTAIN also operates in unsupervised mode (without feedback), forming clusters based on stimulus similarity alone. This dual capability allows it to model both intentional category learning and incidental learning of statistical regularities.