Possible / possible category in deep learning

Question

I am interested in using some of the partially labeled data that I have in the deep study task. I use a fully convolutional approach, rather than fetching patches from the marked areas.

I have masks that describe areas of certain positive examples in the image, but the non-oiled areas in the images are not necessarily negative - they can be positive. Does anyone know how to incorporate this type of class into a deep learning setup?

Triplet / contrast loss seems like it might be a way, but I'm not sure how to place a “fuzzy” or ambiguous negative / positive space.

Answer 1

Try label smoothing as described in section 7.5.1 of the Deep Learning Book :

It can be assumed that for some small constant eps mark y of the training set is valid with a probability of 1 - eps , and otherwise, any of the other possible marks may be correct.

Label smoothing organizes the softmax-based model with output values ​​of k , replacing the hard 0 and 1 classification targets with eps / k and 1 - (k - 1) / k * eps targets, respectively.

Check out my question about introducing label smoothing in Pandas .

Otherwise, if you know for sure that some areas are negative, others are positive, and some are undefined, then you can introduce a third undefined class. I worked with datasets that contained an undefined class that matched patterns that could belong to any of the available classes.

Answer 2

I assume that you are struggling with the segmantation task with a poorly defined background problem (for example, you are not sure that all examples are correctly labeled). I recently encountered a similar problem, and this is what I discovered during my research:

• In the old days, before deep learning and at the beginning of the deep learning era, a common way to deal with this is to smooth out your results with some kind of probabilistic model that takes into account the possibility of noise (you can read about this in the chapter Studying shortcuts from noisy data from of this book , discriminate these probabilistic models from the models used to smooth your labels in relation to the image structure or labels, for example, classic CRFs for two-sided smoothing.

• What we finally used (and worked great) is the idea of ​​Channel Inhibited Softmax from this article . In terms of mathematical properties - this makes your network much more reliable for some objects that are not tagged - because it forces your network to output a much higher positive rating of logits to correctly tagged objects.

Answer 3

You can see this as a problem with semi-surveillance. Use the full, unlabeled dataset to train the structure of the auto encoder bottleneck (or GAN approach). Then this pre-processed model can be adjusted (for example, remove the last layers, add a better layer structure at the end on top of the bottleneck functions) and finit with the marked data.