MLOPS C3W2 video 1 results. Which is actually better?

In the video the reduced dimensionality model is said to be significantly better than the 1000 dimension version. I started wondering is it really so.

The graphs do not show so much overfitting, but the validation accuracy seems to be about the same (0.55) and 6-dim model loss is 1.8 compared to 0.9 in 1000-dim model.

Did I not understand something correctly here? To me the dimensionally reduced model really doesn’t seem much better, otherwise that lower overfitting, that comes from lower train accuracy.

Hi @j-p . In general terms, the idea is that the distance between curves up to epoch ~ 20, is definitely lower in the 6-d model than in the 1000-d model. And that is an indication of better performance somehow. Let me know if this is a property you are able to see or not as a general feature of the curves?