Figure 3:

Transitive inference and realistic behavioral patterns. a: Meta-training loss: Mean performance in the last 10 trials of each episode, for 30000 episodes of meta-training (30 separate runs, all curves smoothed with a boxcar filter of width 10). Notice the two slightly lower curves, indicating two runs settling on the suboptimal solution (see Section 5). b: Test performance: mean performance in the last 10 trials of 2000 test episodes (with the same frozen network but different randomly generated stimuli), for each pair, arranged by “symbolic” (intra-pair) distance. More distant pairs elicit higher performance (symbolic distance effect), as observed in animal and human experiments.