19_dec_2016

Tracking the World State with Recurrent Entity Networks

Questions:

Edgar: Given the inputs, how does the network differentiate the contents and the keywords?

@priba:The keywords must be previously fixed or are learned by the system? For each different question (where, when, how, etc.) the system must be trained? What information do the system have during the training?

@gitabcworld: How the normalization at the final step allows to forget previous information of each block? They explain the information of the memory is contained within a unit sphere and each new memory is decreasing the cosine distance between the original and the updated memory. How is this done in the dynamic memory schema?