• Runtime breakdown
• Analysis 

m params, n tuples = m x n matrix

• Batch deletion (already have).
• Sample tuples with higher influence value in the previous round.
• Cluster tuples and only do gradient on the cluster center.