22:21 < travis-ci> mlpack/mlpack#710 (master - f0675d7 : Ryan Curtin): The build is still failing.

22:21 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/5bc514c122d5...f0675d7789b6

22:21 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/121282997

23:01 < uzipaz> zoq: hi zoq, is there a way to specify a confidence threshold when using feedforward neural nets for classification? Thanks

23:05 < zoq> uzipaz: Hello, at the moment, no, but it's easy to add, we could modifiy the BinaryClassificationLayer class and use a confidence parameter in line 61 instead of 0.5.

23:06 < uzipaz> zoq: thanks for answering, I am trying to plot precision-recall and roc curves for the nerual nets Im using

23:07 < zoq> uzipaz: nice, you could also use the MulticlassClassificationLayer class and do everything you like afterwards.

23:08 < uzipaz> zoq: I just read the code you referred and Im wondering why you guys chose to hardcode the confidence threshold?

23:11 < zoq> uzipaz: There is no reason behind it, if you like you can open a pull request, don't feel obligated, I'll can do that probably tomorrow.

23:17 < uzipaz> zoq: I am looking at MulticlassClassificationLayer class and in the OutputClass function, it just assigns the inputActivations to output, and the inputActivations just represent a probability between 0, 1

23:21 < zoq> uzipaz: Right, so you could use your confidence value, after the OutputClass function is called. But I think, modifying the BinaryClassificationLayer class is the best solution.

23:23 < uzipaz> zoq: Im sry im bit confused in case of 'n' MulticlassClassification, each data sample for which we are predicting the output class, we need to return 'n' probability values as the result?

23:23 < uzipaz> zoq: but incase of binaryclassification, we just return a single probability value for the output class

23:30 < zoq> uzipaz: hm, the BinaryClassificationLayer and the MulticlassClassificationLayer class returns a vector with n values (probabilities), but the BinaryClassificationLayer transforms the output of the n dim input vector, I guess the naming is confusing.

23:35 < uzipaz> zoq: Im concerned about MultiClassClassificationLayer, in outputClass function, each element of argument 'output' should be a list of 'n' probability values, where 'n' is the number of classes in our multiClassClassification problem

23:36 < uzipaz> zoq: also, how does the MulticlassClassificationLayer class knows how many number of classes are there?

23:45 < zoq> uzipaz: In most cases the MulticlassClassificationLayer is the last layer of the network, so the number of classes is the number of elements of the inputActivations vector, which is also the number of units in the last layer.

23:47 < uzipaz> zoq: so in binaryClassClassificationLayer, the number of elements of the inputActivations vector will be just 1?

23:51 < uzipaz> zoq: its just a single probability value for binaryClassClassificationLayer?

23:52 < zoq> uzipaz: The number of elements of the inputActivations depends on the last layer. So if the last layer has 3 units (3 classes) it could be something like [0.3 0.8 0.2]. If you use the MulticlassClassificationLayer you get as output [0.3 0.8 0.6]. If you use the BinaryClassificationLayer you get as output [0 1 1] and if you use the OneHotLayer you get [0 1 0].

23:57 < uzipaz> zoq: its much clear to me now, thank you for explaining, when calling the predict function on a network that has 3 units in last layer and 10 function point 'predictors', the size of argument 'response' will be 3 x 10? where each value is between (0, 1)?

23:59 < zoq> uzipaz: yes, right