Figure 2.

The DNF model’s architecture. The model has two fields, one (A) organized by space and a dimension formed from an MDS solution of Xu and Tenenbaum’s (2007a) cluster trees, the other (B) by label and the same MDS dimension. Each unit (e.g., C) is tuned most strongly to a particular value along its fields’ two dimensions. Objects are represented by peaks of activation (D) in the space-MDS field, that after being weighted by a Gaussian kernel (result E) project ridges (F) in the label-MDS field. External inputs to the model are Gaussian patterns representing labels (G) that project vertical ridges of their own into the label-MDS field, two dimensional Gaussian patterns representing exemplars (H) that drive the formation of the peaks in the space-MDS field (D), and Gaussian patterns representing test objects for generalization (I) that project additional horizontal ridges (J) into the label-MDS field. Following the model through a test trial to Row 2 after 40 time steps, the exemplar peaks have strengthened, projecting activation (K) into a stronger ridge (L), which is almost forming a peak with the overlap of the test object ridge. 40 time steps later in Row 3, the model has built a generalization peak (M). In an alternative trial with a different test item (Row 4), the test item is in between the features of the exemplars, and does not overlap enough to raise a generalization peak.