Are there example implementations of networks that apply constraints across sequences of image classifications where class labels are ordinal numbers? For example, to cause the output of a CNN to monotonically increase across frames, where the number may increase either more or less steeply but only in one direction across the entire sequence, or as another example, to smoothly vary rather than jumping precipitously from frame to frame. In my first example, the output can jump quickly from one frame to the next, as long as only in one direction, whereas in my second example, they can either increase or decrease as long as not too "fast" from one frame to the next as if being passed through a low pass boxcar filter. The first is a monotonicity constraint and the second is a smoothness constraint, but in both examples, the key is for adjacent frames to have an effect on the conclusions for a given frame.
Thank you, Andy