Deep learning 4j vs TensorFlow vs Multilayer Perceptron predictor

[zizoo]

Hi,
Is there a comparison between these three algorthim that are available in Knime.
DL4j crashed with my dataset and I used multilayer perceptron that worked well.
As I use with the later method many layers, can I consider it as equivalent to DL4j or Tensor Flow?

[nemad]

Hello @zizoo,

equivalent in what terms?
If you only have a small dataset, with few features, then one could say the three are somewhat equivalent.
However, if you are looking for more advanced deep learning functionality, there is no way around our Keras integration, as it provides by far the most functionality (including GPU support).

Cheers,

Adrian

[zizoo]

Hi @nemad,
I have a dataset of thousands of samples. I can train them on DL4j or MLP. If I use the same hyperparameters, should I expect similar results?

[nemad]

Hello @zizoo,

this depends on so many factors that I don’t feel comfortable to claim that. Even for the same dataset, same architecture and same framework, different network initializations can result in different results.
I am also not even sure if you actually can specify the same hyperparameters in KNIME’s MLP and DL4j because our RProp MLP Learner uses a specialized version of backpropagation and I don’t know if you can configure that in DL4j.