How do weights changes handles during back-propagation when there are unknown labels


I have a question about how weights are updated during back-propagation for some of my samples that have unknown labels (please note, unknown, not missing). The reason they are unknown is because this is genomic data and to generate these data would take 8 years of lab work! Nevertheless I have genomic data for samples that have multiple-labels, sex age organ etc. this is a multi-class multi-label problem.

For most classes, ALL labels are complete. For two or three classes, there are unknown labels. an example would be the developmental stage of samples at age x, the developmental stage of sample at age y are known. the developmental stage of samples at age Z are unknown! (generating this data is what would take most time)... I would therefore like to include all this data during training as it is indispensable. I would like to generate the sigmoid probability and assign unknown label 'Z' as belonging to developmental stage 0 or 1 (known classes) based on a threshold (say >= 0.5)... When one-hot encoding the unknown labels simply have no ground truth, 0 for class developing and 0 for not-developing as follows (example of 3 samples shown for class in question):

  [[1., 0.
   [0., 1.
   [0., 0.  ......]]

the first row is known sample 1, second is known sample 2 and 3rd is unknown, and therefore has no ground truth. It is this sample i would like to assign a label of known class 1 or 2 based on the 'highest probability'.. based on reading and discussions, this is the direction i will be taking for this task, as it can be validated in the lab later... so the approach is, include in training and see what the network 'thinks' it is.

My question is: How does back-propagation handle these known and unknown samples with respect to weight updates.

I should note i have trained the network with ~90% validation performance. for all classes for which is there is complete data, the predictions are great. and the same for classes for which there is unknown data. It can accurately classify the samples for which there is known developmental stages... and it does assign a probability value to those samples that have the 'unknown' label (0,0), so i would really like to know how back-prop is handling these samples for the classes where there are unknown ground truth labels.

thank you!


Posted 2020-03-27T21:56:43.117

Reputation: 41

No answers