San Francisco - September 9, 2015
Jason Brown on Epidemic Algorithms for Replicated Database Maintenance
Stephen Tu on Random features for large-scale kernel machines
"While published in 1987, 'Epidemic Algorithms for Replicated Database Maintenance' describes an early distributed system that uses gossip replication and eventual consistency (the latter uncannily like modern NoSQL databases).
The subject of gossip branched into some interesting directions since the '80s, I'll primarily concentrate on the Demers paper, give a brief survey of the gossip literature, and tie those papers into some modern, well-known systems (namely, Cassandra and Riak)."
Jason Brown is a Senior Software Engineer at Apple, as well as being an Apache Cassandra Committer. He holds a Master’s degree in Music Composition is searching for time to write a second string quartet.
From Stephen Tu:
Kernel methods in machine learning are a popular tool used to express richer function classes. These methods work via access to a kernel function, which can be thought of as a measure of similarity between two vectors. The standard method of optimization via kernels requires solving a program with decision variable the size of the number of data points. As dataset sizes grow, this becomes prohibitive.
This paper addresses this issue in a surprisingly practical way. Specifically, the authors show that by using a multiple of d random projections based on the Fourier transform of the kernel function (where d is the ambient dimension), one can accurately approximate the kernel function. This allows practitioners to use fast linear solvers and retain the benefits of more expressive function classes.
Stephen is a graduate student in EECS at UC Berkeley.
The San Francisco Chapter would like to give special thanks to Fastly for sponsoring the September meetup.