Why we need to shuffle the samples in SGD ?

[skwong]

In hw5, Q8 Q9. I've done the SGD, then I tried not to shuffle the 100 samples, instead, calculate from 1 to 100 in each epoch, and they come up with similar result.

[yaser]

Originally Posted by skwong

In hw5, Q8 Q9. I've done the SGD, then I tried not to shuffle the 100 samples, instead, calculate from 1 to 100 in each epoch, and they come up with similar result.

The shuffle is done anew for every epoch to get the benefit of randomness. For some data sets randomness doesn't make that much of a difference, but for some it does.