Questions for 6.S897 lecture 18 (Machine Learning, 11/19).
Email your answers to 6.s897staff@gmail.com.

1) What features does the Parameter Server provide over a conventional key-value store?

2) Why were vector clocks used in the Parameter Server, if many updates to the model are commutative operations like addition?

3) How does the data partitioning across machines in Adam reduce communication?