Figure 6

Interpolation of the test data set feature vector ‘Jaguar’. Three semantic features (dangerous, having a fur and having lungs) are missing, i.e. are replaced by the value \(-1\). The three networks trained with different discount factors \(\gamma = (0.3,0.7,1.0)\) infer the missing features. Binary semantic features are inferred well in all cases. The ‘dangerous’ feature is badly predicted for large discount factors \(\gamma = (0.7,1.0)\). In contrast, in case of the lower discount factor \(\gamma = 0.3\), it is predicted well.