There is an interesting comparison in this post that compares Avro, Protobuf and Thrift of binary messages sizes and how well the protocol supports schema evolution. Another interesting data transfer protocol is Parquet, which is optimized for column- oriented data.
Today I was at a Computation meets Data Science Conference, organised by Wolfram Research and the CQF. There were some interesting talks. The ones I enjoyed the most used Mathematica to analyse data in real time in interesting ways. It looks like Mathematica has good support for building neural networks now. I was impressed at how quickly Jon Macloone from Wolfram was able to get some quite useful neural network models up and running. Jon made the point that for some problems you are able to get results really quickly with neural nets, and others it’s really hard to get good results, and it’s not obvious which problems are which.