Nuts and Bolts of Applying Deep Learning (Andrew Ng)
Why Cross Validated Stack Exchange recommends it
Stats Stack Exchange (Cross Validated) users link it as a citable source for practical rules on train/dev/test splits and on reading train versus test error gaps.
Recommended by 1 people across 1 different months, 2017 to 2017.
What it is
Andrew Ng's talk from the 2016 Deep Learning School in the Bay Area, posted by Lex Fridman as a clip of the full livestream. It gives practical advice on using deep learning in real projects, including how to split data into train, dev and test sets, how to read bias and variance, and how to decide what to fix next.
What people said
Recommends ednaMode · 2017-07 · Cross Validated Stack Exchange · on “Train vs Test Error Gap and its relationship to Overfitting : Reconciling conflicting advice”
There seems to be conflicting advice out there about how to handle comparing train vs test error, particularly when there is a gap between the two. There seem to be two schools of thought that to me, seem to conflict. I'm looking to understand how to reconcile the two (or understand what I'm missing here). Thought #1: A gap between train and test set performance alone does not indicate overfitting First, (also discussed here: How can training and testing error comparisons be indicative of overfitting? [LINK] ), the idea that a difference between train and test set alone cannot indicate overfitting. This agrees with my practical experience with, for example, ensemble tree methods, where even
Cites MAA · 2017-08 · Cross Validated Stack Exchange · on “Should the validation set come from the training set or the testing set?”
Nuts and bolts of applying deep learning [this video] Never mind. Andrew Ng states here that the validation set and the testing set should come from the same distribution (42:10). But what does he know? "Splitting your data. In most deep learning problems, train and test come from different distributions. For example, suppose you are working on implementing an AI powered rearview mirror and have gathered 2 chunks of data: the first, larger chunk comes from many places (could be partly bought, and partly crowdsourced) and the second, much smaller chunk is actual car data. In this case, splitting the data into train/dev/test can be tricky. One might be tempted to carve the dev set out of the traini
Cites Krrr · 2017-09 · Cross Validated Stack Exchange · on “Why is everything involving machine learning called AI lately?”
I disagree that AI is just a "fancy" name for ML used for marketing, as some of the comments seem to say. The "immediate" aims of AI is to computationally solve problems that humans (or some other biological intelligent agent) can solve naturally in short time (something like a few hundred milliseconds). The examples include, face recognition, natural language understanding etc. I think Andrew Ng says it somewhere in this video [this video] . Also see this article [LINK] that tries to clarify the difference.