Skip to main content
Fig. 2 | Visual Computing for Industry, Biomedicine, and Art

Fig. 2

From: Visual analytics tool for the interpretation of hidden states in recurrent neural networks

Fig. 2

Detailed view for one selected sequence from the IMDB dataset. Different information and interactive visualizations are shown on panels (A) to (D). (A): The sequence ID and information about the classification; (B): The input sequence with each word colored according to its expected prediction (EP), i.e., what output the model would produce if that word was the last in the sequence; (C): A heatmap matrix displaying how the model’s prediction evolves during the sequence processing for each class (top) and overall (bottom); (D): A histogram of the Euclidean distances between the hidden states of the input sequence; (E): The center highlights in the projection the hidden states produced by the sequence, giving insight into how the hidden state evolves over the sequence processing. It is possible to define a threshold for distances to show words in the projection resulting from a larger change in the hidden state; this threshold is also visible as a horizontal line in the histogram; (F): On the right side, the currently selected sequence is highlighted in a list. In the example used for these visualizations, we note how the model first believed the sequence to be classified as positive during the first few time steps of the input sequence. However, as it obtained more information about the input, it changed its output to a negative value. Underlying data source: IMDB as available in Keras [13]

Back to article page