Thesis

Efficient model fitting of artificial neural networks.

Explaining

This research's purpose is to automatize the process of finding the right model by filtering the wrong parameters those only describes the training dataset specific features and not valid in the general model.

It is a special generation based regularisation that observes the training process and like a human would do with a heuristics it evaluates the model parameters, determining the general correlations on the dataset in the specific model.

Main strenghts of the algorithm is the O(n) scaling ('n' is the parameters's number), so it doesn't add more complexity on the calculation on neural networks. Weakness, it needs more improvement for handling bigger dataset.

Short conclusion

This method is capable of decreasing the overfitting in an automatic way.

Document

The thesis is written in Hungarian. A beta version is available: https://tdk.bme.hu/VIK/DownloadPaper/Mestergeges-neuralis-halok-modellenek-javitasa The master thesis will be available for public in 2019. Till that time it is owned by Budapest University of Science and Economics.

In short of the full publication

• Background
• Specifications/implementation
• Measurements/proof, comparisons with other regualrization method (DropOut (In many case my algorithm beat it.))
• Literature

Used technology

• Tensorflow
• Python (for fast prototyping)

Algorithm limits/conditions

• There must be some kind of overfitting between the train and validation dataset (that is what it decrease automatic way.)
• The arithmetic used in the neural network is limited (It can be extended.) ..* e.x. recurrence node needs some work to be handled properly
• Batch based training isn't adequate

Results

The prototype was able to decrease overfitting on test datasets by order of magnitude: [results will come here]