Communication requirement for distributed statistical machine learning with application in waveform cognition

Distributed learning is an effective approach to mitigate the data communications in machine learning when the data is stored in a distributed manner, particularly in the era of big data. In the distributed learning procedure, learners can send intermediate computation results instead of raw data, thus reducing the communication cost. In this paper, the communication requirement for distributed learning is studied in the scenario of multiple data storage nodes having the capability of learning and a fusion center. Lower bounds for communications are derived based on VC-entropy of modeling in the machine learning. Numerical results are provided to show the communication requirement for typical learning problems.