Is it possible to know if relation R, is true about two sets of input using neural networks I have just seen an introduction video to neural networks, and I was wondering, are neural networks only useful for when you want to say that "A based on it's features matrix, belongs to category C", or can it also be used to detect relations as well.
What I mean is, is it possible to say that "if you (the computer) see these numbers for A [1 (male), 4 (blood type B+), 78 (weight), 23 (age)] and you see this type of matrix for B [1 (male), 5 (blood type B-), 82 (weight), 25 (age)], then 'learn' that they have the relationship R (1 = are brothers), else they are not related (0) and if I give you new sets in the future, decide if they are related or not based on what you have learned"? In other words, is it possible to detect relationship between two sets of input samples, in a manner that the relation X is true only about these people from group A and these people from group B; based on their matrices?
If neural networks are only for classification of only one group, what machine learning method can be used to train the computer (or make a model) to detect if relationship R is true about two streams of input or not? Because even clustering that seems a bit close, decides if one items belongs in a category or not.
I know that probably these two groups can be combined into one huge group, and be arranged randomly, to see if they are brothers or not, but other than that method, which seems pretty expensive, are there any other methods?
 A: The thing to recognize here is that although this looks different than a prototypical classification problem, it is ultimately the same.  Instead of having data on various individuals, you have sets of data on pairs of individuals.  The classes to which you want to assign your inputs aren't, say, spam vs. good, but brothers vs. not-brothers.  You would need think carefully about how you design the neural network if you want optimal performance, and not just use a default fully-connected feed-forward network.  For example, you might want a convolutional NN with specific receptive fields.  You might also want to do some preprocessing of your data and feed that into the network for training instead of using the raw data.  For example, you might want to determine similarities between each pair of measurements (e.g., heights) according to various measures of similarity, and use one set of similarities as your inputs instead of two sets of raw measurements.  This would certainly be a more complex project than the standard textbook examples, but the point is that it is ultimately similar.  
