Telcos have vast amounts of data about user behaviour, but it is often extremely highly dimensional and sparsely populated, making it hard to learn from using conventional ML techniques. By compressing this data into a smaller space, we can build simple downstream models that leverage subtle features in the data - like predicting gender or churn - without handcrafted features.
The Primary Objective is to learn about implementing and using an auto-encoder (and their utility). The secondary objective is to learn about the practicalities of actually implementing an auto-encoder in MxNet.
This talk will not contain a lot of math. It will discuss implementation, loss functions, and the idea behind KL divergence.
The talk will first lay out prerequisite knowledge (basics of deep learning) and then introduce auto-encoders. It’ll then introduce the context of their use in this talk. It will successively introduce more complex auto-encoders applied to a real world problem, in particular concentrating on choice of loss function, then multiple tasks, and finally utilising KL divergence for sparsity.
This talk will teach you about: