Here, I tend to give my personal explanation of ELK concept and task based on the terms that have already emerged (Still in progress.). 


The original statement of ELK: Given input samples set X (we are interested in), targeted model M, and task (goal) G, could we obtain a reporter or direct translator to honestly tell us the internal mechanism of M on G given X?

Translating into the existing terms: obtaining faithful model distillation on specific task or property. 

faithful == honest, reporter == surrogate model (distilled or not)