Gossip-Based Machine Learning in Fully Distributed Environments
István Hegedűs · 2017
FIGURES 5.7 The effect of network size (with sampling rate 1/∆). . . . . . . . . .5.8 The effect of the churn. . . . . . . . . . . . . . . . . . . . . . . . . . .5.9 Model age histograms over online nodes for long session lengths. .5.10 Model age histograms over online nodes for short session lengths. .6.1 Convergence on the real data sets.Error is based on cosine similarity.In the scaled version of GRADSVD the number of iterations is multiplied by log 10 m (see text). . . . . . . . . . . . . . . . . . . . . .6.2 Convergence on the real data sets.Error is based on the Frobenius norm.Horizontal dashed lines in top-down order show the FNORM value for the optimal rank-i approximations for i = 1, . . ., k. 6.3 Results on synthetic data sets using networks of different dimensions.We set k = 1, and all the matrices had a rank of 16. . . . . . .6.4 Results when only the 50/33% randomly sampled instances were used from the data set. . . . . . . . . . . . . . . . . . . . . . . . . . .6.5 Results in different failure scenarios using a 1024 × 1024 synthetic matrix with a rank of 16.We set k = 1. . . . . . . . . . . . . . . . . .