Error correcting codes for distributed non-linear function computation

Open Access

Author:: Tapha, Thitiwat
Area of Honors:: Electrical Engineering
Degree:: Bachelor of Science
Document Type:: Thesis
Thesis Supervisors:: Viveck Ramesh Cadambe, Thesis Supervisor
Julio Urbina, Thesis Honors Advisor
Keywords:: Communication system
optimization
gradient descent
alternating optimization
logistic regression
loss function
matrix-multiplication
Abstract:: We provide novel coded computational strategies for distributed non-linear function computation. We utilize the idea of previous studies on coded matrix-multiplication computation and extend our study to achieve optimal error in non-linear distributed logistic regression. We consider the system of P computation nodes where each node receives two one-dimension input matrices. Our computation system reduces the error from stragglers via encoding and decoding. Our experiment with a random set of the number of worker nodes, threshold, and number of inputs has yielded conclusions on the effect of parameters on the loss function. Moreover, we develop an optimization framework based on alternating minimization that enables the discovery of new codes to achieve optimal coefficients to minimize the error in non-linear functions.