"Defining deep learning"

[davharris]

I think that this section doesn't quite accomplish what it should, in terms of highlighting deep learning's distinctive features and its relationship to "shallow" approaches that readers might be familiar with. I have a small outline below.

I've tried rewriting it a few times, but I'm feeling stuck. If someone agrees that these would be improvements and wants to try them out, go for it. Otherwise I'll try again in a few days.

• Should mention the key features of deep learning, compared to other methods:
  • composition simple functions in "layers"
  • differentiability/backpropagation (in most cases)
  • automated feature engineering
• Should clarify that deep learning does many of the same jobs as more familiar approaches, such as clustering, regression/classification. It just has the potential to do them better because (with sufficiently large data sets) these models can learn which features to use.
• This is very minor, but I think that the current wording implies that autoencoders are "new"

[agitter]

At one point we had a version of the intro that included some of these ideas, but in a fairly technical way that may not have been useful for our audience. In #382 @cgreene worked to merge two versions of the intro, and we stripped out the more technical parts.

I agree that the points above will be important for someone new to deep learning to understand some of the excitement that surrounds it. If you do write something, I suggest trying to keep it to a single paragraph. It could follow your existing outline closely.

[cgreene]

I agree - we lost this & it would be really nice to have.