Most of the people is trying to answer question with a neural network. However, has anyone came up with some thoughts about how to make neural network ask questions, instead of answer questions? For example, if a CNN can decide which category an object belongs to, than can it ask some question to help the the classification?
-
$\begingroup$ I think that doing this qualifies as a second order question. link I feel like I was approach doing this for coplanarity excursion metrics using Eureqa, but it took comprising the question in a strongly atypical way. It was a substantially bigger, harder question. Metrics on metrics, or meta-metrics learning had to be invoked. Specificity helps simplify the question. Specify the domain and bounds of the question. $\endgroup$– EngrStudentSep 12, 2016 at 11:07
3 Answers
Maybe neural networks are not the best tool for this.
It seems to me that an equivalent of the your notion of 'a question to help the classification' would be to use Machine Learning (ML) to obtain a human-readable ruleset which performs the classification. The idea is that, if you follow an applicable chain of rules all the way through to the end, you have a classifier, if you stop before that, you have an indicator of which features of the input give more coarse-grained classifications, which can be seen as a progressively detailed sequence of questions that 'help the classification'.
More detail on various options for using ML to create rulesets can be found in my answer to this question.
One solution to this could involve a fusion of a decision tree and ANN for a multilevel classification.
A decision tree can help with predicting the possible category of the instance to classify. Then, the ANN at the leaves of the tree can produce the final classification.
For example, in image recognition, the tree can decide what category of object to identify (eg., landscape, people, vehicles, etc.) and the ANN for the appropriate type can predict exactly what object it is. In vehicles, for example, car, bus, bike, etc.
Great question. Today AI systems works in "one burst" mode. Get one input and generate one output. Our brains are not working like that.
First step is to learn network how to communicate with it's "helper", so network instead of result generate question and cycle will repeat until network find result.
Network must be recurrent for inner state needed between question/answer cycles.