Tweaking a CNN for large number of input channels


I am using a CNN for function approximation using geospatial data. The input of the function I am trying to approximate consists of all the spatial distances between N location on a grid and all the other points in the grid.

As of now I implemented a CNN that takes an "image" as input. The image has N channels, one for each location of interest. Each i-th channel is a matrix representing my grid, where the pixel values are the distance between each point in the grid and the i-th location of interest. The labels are the N values computed via the actual function I want to approximate. N can be up to 100.

Here an example input of the first layer:

enter image description here

So far I could see the train and validation loss go down, but since it is a bit of a unusual application for a CNN (to my knowledge the input channels are at most 3, RGB) I was wondering:

  • does this many-channel-input approach have any pitfalls?
  • will I be able to obtain a good accuracy or are there any hard limits I am not aware of?
  • are there any other similar application in literature?


Posted 2019-08-12T09:05:22.343

Reputation: 121

It is possible to train a CNN with multiple channels, for example i've seen Wavenet like structures applied to traffic datasets, where they have multiple channels such as: average speed, lane occupancy, number of cars etc. The only downside that I could see is the large parameter space that it might require. – razvanc92 – 2019-09-12T13:30:49.833



As far as I know, more than 3 channel is perfectly fine, since, 3 channels are what we use for images and that's enough since we can only see this many colors, but I don't see why more than that wouldn't work

Your 2nd question is like asking whether or not you will be good at a sport... Just try it

For your 3rd question, I've never seen any language AI using CNN instead they all use RNN, not sure if that's what you meant though


Posted 2019-08-12T09:05:22.343

Reputation: 39