How to get 68% accuracy using an MLP?

[Phoemuex]

Thanks for providing this nice dataset!

In the README file (and also in the paper at arXiv), it is stated that one can obtain a 68% (test(?)) accuracy using an MLP.

However, the output in the relevant notebook included in the repository seems to indicate a test accuracy of ~65.7%.

Moreover, in my own tests with the same architecture that is used in the relevant notebook (i.e., two hidden layers with 100 neurons each), I only managed to achieve around 62% accuracy.

Or were the 68% obtained with a different network architecture (number of layers, neurons per layer, regularization, etc.) than what is used in the notebook that is part of the repository?

Any hints would be appreciated.