This is the second part of Adrian Colyer's summary started yesterday of a PhD thesis by Heidi Howard on consensus algorithms. In this post he describes some simple revisions to the classic Paxos algorithm, then looks at a significant proposal: Multi-Paxos, which is an optimisation of Classic Paxos for consensus over a sequence of values. Again, the summary should be widely accessible if followed carefully step by step. Note that while all of this may seem pretty far removed from online learning, a lot of the thinking here underlies how decentralized learning needs to be designed in order to be robust and reliable (and hence to avoid the failures of centralized systems).
Today: 1 Total: 1099 [Direct link] [Share]
