1: \begin{abstract}

2: \Gls{FL} achieves collaborative learning without the need for data sharing, thus preventing privacy leakage.

3: To extend \gls{FL} into a fully decentralized algorithm, researchers have applied distributed optimization algorithms to \gls{FL}

4: by considering \gls{ML} tasks as parameter optimization problems.

5: Conversely, the \gls{CMFD} proposed in the authors' previous work

6: makes \gls{NN} models get close with others in a function space rather than in a parameter space.

7: Hence, this study solves two unresolved challenges of \gls{CMFD}:

8: (1) communication cost reduction and (2) visualization of model convergence.

9: Based on a proposed \gls{DCCR}, the amount of data transferred in a network is reduced; however, with a slight degradation in the prediction accuracy.

10: In addition, a technique for visualizing the distance between the \gls{NN} models in a function space is also proposed.

11: The technique applies a dimensionality reduction technique by approximating infinite-dimensional functions as numerical vectors

12: to visualize the trajectory of how the models change by the distributed learning algorithm.

13: \end{abstract}

14: