Deep Coaching

Inspiration

There is an interesting and a fundamental question in machine learning: how do we make use of the privileged knowledge which is only known in the training samples but unknown to the test ones? We try to address this problem in a geoscientific scenario.

What it does

Here we compared the performance of learning with and without coaching variables, and of deep learning and linear model, in predicting temperature from METAR Data. We find that deep learning with coaching variables performs best.

How we built it

We used METAR Data from NWS accessible through THREDDS Data Server. We used R to process the data and build the model. The neural network is written using kerasR library.

Challenges we ran into

It took long time to find and process the data. Besides, we tried some models which do not work.

Accomplishments that we're proud of

We find that neural networks do perform better and coaching variables do make a difference.

What we learned

We get a better understanding of how to interpret data especially when it is not so clean at the beginning.

What's next for Deep Coaching

We may try to predict other important variables in the data such as the location. This would be more challenging because it would involve larger amount of data and predict location given weather might be useful in some cases. This has been attempted but not finished yet.

Built With

• r