Assessment of generalization performance of top-down modulation using stream II, IV, V. System-level models were trained on stream I, II, III (baseline) and on stream III, both times using blocking. Results were very similar in nature on streams I, III (not shown). Solid green curves appearance-based classifier without modulation. Solid red curves appearance-based classifier using attentional modulation trained on streams I, II, III. Dashed green curves appearance-based classifier using attentional modulation trained on stream III. As can be seen from the plots, training system-level models only on stream III do not affect performance significantly in any direction