Fig. 7
From: Visual analytics tool for the interpretation of hidden states in recurrent neural networks
Examples of correctly classified sequences in a model trained for multi-class classification. The projection provides insight into the class distribution within the hidden state space. The heatmap matrix supports the analysis by displaying the evolution of the EP over sequence processing. Hence, the user can better identify at which time steps the model was able to distinguish between classes. In the visualizations on the left, the model is certain for the classification of earns (yellow). In the visualizations in the middle, the model is uncertain between earns (yellow) and acquisition (violet), and both classes have a similar EP while processing the sequence. In the visualizations on the right, the model is uncertain among all five classes and the EP for money-fx (blue) is only slightly larger compared to the other classes. Similar cases are likely to be misclassified. Underlying data source: Reuters [25]