How to Keep Consistence in Asynchronous Stochastic Gradient Descent

[bigheiniu]

hi, in the paper it has said that asynchronous update will not make conflict, but when I want to add Polynomial Kernel Function in calculating the probability of observing an edge in low dimension, multithread went not correct, but one thread works well.

So do you have suggestions on keeping parameters consistency, currently I just want to imitate a parameter server to keep the data consistency.

To further accelerate the training process, we adopt the asynchronous stochastic gradient descent, which is very ef- ficient and effective on sparse graphs [19]. The reason is that when different threads sample different edges for model updating, as the graph is very sparse, the vertices of the sampled edges in different threads seldom overlap