Fig. 4

Subset performance optimization through train-time. A: Shown is the improvement of the models trained on the 50 k subset when increasing the training epochs from 50 to 1000 epochs and comparing to the performance of models trained on the full data set. The three architectures illustrated here were chosen because they were the best performing architectures in the experiments illustrated in Figs. 2, 3. Architectures are all GRUs without attention and model abbreviations follow the scheme “latent size—hidden size—number of layers”. Shown are Mean Similarity (gray) and Full Reconstruction (blue) on the test split. The metrics are presented as the mean of three seeds with error bars indicating the highest and lowest value. B: Energy consumption of the three models trained on the full set for 50 epochs and the 50 k subset for 1000 epochs, which achieved comparable reconstruction performance on the test set. *The average energy consumption of using a washing machine in Europe is estimated at 120 kWh per year [24], i.e., 10 kWh per month