How to use non-image input for conv-neural network in tensorflow?

Question

I have a data set ([n][13]) like :

 1)      -3.3 -15.0 41 1026.3 16.1 0 25.9 37.0 0 0 0 280 0

 2)      -3.9 -13.9 46 1028.0 16.1 0 20.4 0 0 0 0 280 0

 3)      -3.9 -13.3 49 1028.8 16.1 0 22.2 0 0 0 0 270 0

 4)      -4.4 -12.2 55 1029.5 16.1 0 11.1 0 0 0 0 260 0

...

 n)      -1.1 -10.6 49 1030.0 16.1 0 14.8 0 0 0 0 280 0

Using this data set, I want a create convolutional neural network in Tensorflow and make a prediction.

I know convolutional neural network actually for images but I saw some example with non-image inputs.

Can I create a convolutional neural network with using this data and How can I create? Can you give me any clue or tutorial or source?

Answer 1

This does not look like data that is suited for a convnet - convnets assume that it makes sense to share the weights locally, so they make sense when there's some form of locality in data. So it makes sense for other things than image if there is a connection from a feature to "nearby" features - for example in time series, or audio, where features that are adjacent happened at nearby time point. Your data looks like the columns are unconnected (even on different scales).

Your data is also really low-dimensional so you can just go with fully connected layers, if you want to go for a deep learning approach - I would try out other approaches on your data as well (boosting & random forests).

Answer 2

Actually, I was trying to training deep learning model from low dimensional data, exactly 8 features. I designed 11 layer-neural networks. However, it does not show quite different performance, comparing to previous model trained by Naive Bayes algorithm. Like you, I'd like to lean neural network algorithm through hands on experiments. Anyway, one thing I want to tell you is you should normalize your data to have same scaling. If not, Neural network won't be converged.