Performance results are based on testing as of June 1, 2021.

The same automatically preprocessed dataset, without feature engineering performed, was used for each algorithm within the same Kaggle problem.  

Neuton’s results were achieved with DEFAULT parameters without manual tuning.

For TensorFlow, the most efficient and accurate of 7 pre-prepared architectures was chosen, having from one to 3 layers, and from 8 to 256 neurons in each. In some architectures, dropout layers were also used. The architecture was chosen based on the average validation value over 3-fold validation.

For each non-neural network algorithm, the hyperparameter grid contained up to 216 combinations of parameters, from which up to 50 combinations were selected for random enumeration, based on 3-fold validation followed by training on the full dataset.