In real life when do you stop backward propagation?

Course Q&A Math for Machine Learning Calculus for Machine Learning and Data Science
1

in both wk3 labs, the explanation for stopping was, essentially, “the cost function does not seem to change very much any more”.

in the case of lab 1, the cost function started giving the same value very soon. However in one of the visualizations earlier (wk 1 or 2), there was one case where the propagation bounced between the “left” and “right” of the lowest point seemingly perpetually, which would be consistent with the cost function converging to the same value.

and in lab2, if I increase the # iterations from 100 to 1,000 to 10,000 and to 100,000, the line clearly gets flatter and flatter even if only by a little bit

can someone explain or give an example in real life when is propagation stopped? and in which course will this topic be discussed? thanks

2

Any sort of fixed limit is going to require making assumptions about the data set and how well the model fits.

The rule of thumb is “stop iterating when you get a solution that works good-enough for whatever problem you’re solving”.

If the solution is oscillating around the minimum, the learning rate is too high.

1 Like
3

Maybe the other thing worth saying is that “good enough” will be determined in terms of prediction accuracy, not the actual J cost value. The value of J by itself is not really very meaningful, but it’s useful as a proxy for whether convergence is working or not. As Tom says, if the cost is oscillating, then that gives you “actionable” information: you need to adjust the learning rate.

1 Like
4

thank you TMosh and paullinpaloalto. much appreciated

1 Like