Answer to question 1: The answer is yes, but not very comfortably. That means a very deep feedforward network such as deep MLP or DBN can classify them with too many iterations.
However, also to speak frankly, MLP is the weakest deep architecture and is not ideal for very high dimensions like this. Moreover, DL4J has deprecated DBN since the DL4J 1.0.0-alpha release. Finally, I would still like to show an MLP network config just in case you want to try it:
// Create network configuration and conduct network trainingMultiLayerConfiguration MLPconf = new NeuralNetConfiguration.Builder().seed(seed) .optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT) .updater(new Adam(0.001)).weightInit(WeightInit.XAVIER).list() ...
