How do I configure a network to accept an arbitrary object as the network input?

As most of the examples provided use images as their input, I was wondering if there is logic in place to allow for an object or arbitrary sized non image to be accepted as a network input? I have done work with other fixed sized data inputs into models and see that differing images sizes can now be accepted in some cases.

Answer 1

[Walter Roberson]

There is no logic in place to allow arbitrary objects to be accepted as inputs for neural networks. Mathematical calculations are done on the inputs, but arbitrary objects do not support mathematical calculations.

Neural Networks require numeric vectors of "features", but the underlying tools do not care whether the features are extracted from "images" or audio files or just plain "data".

Some of the interfaces such as Convolutional Neural Networks expect images because image processing is what they are optimized for.

If you have a particular interface in mind, then possibly someone could estimate the effort to change it to use a different form of input.

[Stephen Cronin]

Thank you for your quick reply. My intended use case is in point clouds, initially 3 dimensional, but eventually N dimensional. I know I could treat this as just a list of points, I have been developing operations optimized for this data and my thoughts on the object was as a datatype specific operation, however what you have said on that makes sense.

[Walter Roberson]

You could create a simple function that extracted the location information from the point cloud and ran the NN on it.