I’m hoping to use this course to better understand natural systems.
Is there a way for the neural network to estimate the importance of each of the layer zero datasets in calculating the answer?
If you remove the dataset or value and look for a drop in accuracy, I guess that is one way. Have you already created a program to evaluate data sources in the overall calculation, to see if they are irrelevant to predicting the answer or duplicates of variation from other variables?
Thank you for any direction or ideas.
Ed 
Do you want to do feature selection using neural networks?
1 Like
After a fashion, yes.
I am looking into both into throwing a bunch of data types into the neural network to get the most accurate response, but also to see which data types are not important. In addition, I am hoping to use neural networks to determine spatial/temporal distribution of relevant data in three or four dimensions.
Hey @ed_carter,
By different data-types, are you referring to the fact, that you want to feed images, audio, text, etc into your neural network, and find out the one or the ones which is/are the most important, or am I missing something?
Additionally, when you say that you want the neural networks to determine the spatial/temporal distribution of relevant data, are you referring to performing clustering in either spatial or temporal domain using neural networks, or once again, am I missing something?
Regards,
Elemento
Thanks for the question, Elemento.
Yes, I am looking for which data are the most relevant.
Raster radar rainfall is like an image, as are land use, satellite thermal images, cloud reflectance , antecedent drought conditions, minimum daily temperature, max-min temp difference.
Streamflow or water elevation at a river gage or well is like an audio signal. Then there is 3D stratigraphy , spectral data, etc.
I am not sure how clustering works, but if it is determining variations at which pixels are relevant to determining change in a variable, then yes. If clustering included lag analysis and weeds out autocorrelation to predict target variable changes in both direction and amplitude through time, or under which seasons and conditions then definitely yes.
That’s a lot of different types of data. Generally speaking, a given neural network processes one type of input. So, e.g., if you want to process raster radar rainfall image data and satellite thermal images, those are probably two separate systems. Well, unless you can figure out some technique for creating a blended image with both data included. E.g. you could have a raster image with lots of channels per pixel.
On the more general question of detecting autocorrelation, there are mathematical techniques for doing that. Have a look at Principal Component Analysis (PCA), which uses the mathematical technique Singular Value Decomposition (SVD) as its underpinning. Prof Ng covers that in his original Stanford Machine Learning course. But just googling PCA and SVD should find plenty of useful hits.
2 Likes