Hedging your bets, hierarchical_classifier, DARTS [understanding, questions]

[taras sereda]

Hi all.

I've spend some time trying to understand all the details about hedging project and usage of hierarchical classifiers.

1. I've read the paper http://www.image-net.org/projects/hedging/

2. I've got running examples of DARTS algorithm with models and features provided.

But I still have some concerns and questions about particular usage of image.net.pickle  web_demo.py 

Source code:

https://github.com/BVLC/caffe/blob/master/examples/web_demo/app.py

image.net.pickle has the following structure.  

the format is 

key , len()

 [('probmat', 1860),
 ('words', 1860),
 ('preferences', 1860),
 ('idmapping', 1000),
 ('infogain', 1860)]

words - are from synset of 1000 classes (also including internal nodes)

preferences - probably indicates the depth of each node in semantic tree structure. (since exactly 1000 of items have value of 1)
idmapping - probably holds order of classes as in trained model
infogain - measure of information gain on each node. which depends only on depth of the node. Am i right?


and the last one is probmat which I don't understand how to compute. Can somebody give me a link where I can find more info about this parameter?


From my understanding of Info_gain measurement :
info gain is a quotient of prior distribution to posterior distribution.
or a measurement which describes how entropy changes traversing tree from one level no another.

        # A bias to prefer children nodes in single-chain paths
        # I am setting the value to 0.1 as a quick, simple model.
        # We could use better psychological models here...
        self.bet['infogain'] -= np.array(self.bet['preferences']) * 0.1

while initialisation self.bet['infogain'] immediately reduced by value np.array(self.bet['preferences']) * 0.1?
What exactly it gives us? Is it for describing a fact of a kind of uniform prior distribution?

till now self.bet['infogain'] is a kind of starting value which will be changed in this way for particular predict values:

            # Compute expected information gain
            expected_infogain = np.dot(
                self.bet['probmat'], scores[self.bet['idmapping']])
            expected_infogain *= self.bet['infogain']

self.bet['probmat'].shape
>>(1860, 1000)
scores[self.bet['idmapping']]
>>(1000,)

np.dot function produces matrix with dimensions (1860,1) which in turn will be multiplied element-wise with self.bet['infogain'].
My question is. Where bet['probmat'] comes from? Since in original matlab code I haven't seen such computations.
Thanks in advance.



Regards. Taras.