Talk:Neural Architecture Search

How the second paper builds off the first was really clear with ProxylessNAS improving on the memory required for DARTS! The results were also useful for DARTS in showing the reduction in GPU days.

• I think more related pages could be added though such as convolutional networks and recurrent networks since the papers are looking for convolutional architectures and recurrent architectures.

First paper

Content

• It might be clearer to say "The cell contains two input nodes and a single output node" instead of "The cell comprises".

Prelims

• I think it should say "In convolutional cells, the input nodes of the cell are defined as" rather than "the output nodes".

Method

• I think it would also help to clarify that k is the top-k strongest operations.

Results

• A few sentences to elaborate on the last two tables might be nice to highlight any key findings or state outside of the caption that the second one is for recurrent architecture rather than convolutional architecture.

Second Paper

• It mentions one-shot architecture. It could help to state what one-shot architecture is.

This is an interesting direction! Here is my critique:

• DARTS is not presented in a way that is super clear to me. ProxylessNAS looks much more comprehensive. DARTS should have the same section outline as ProxylessNAS.

• It would be nice to mention what is the same and what is different between the two methods, as it is not clear to me.

The performance of these models does seem pretty good.

• I would've really liked a pseudo code for your second method rather than your first. The first method's pseudo code is not useful.

• It seems there are no experiments comparing the second method with your first memory-wise. I think if it exists in the original work adding it to the page will definitely be more convincing to readers!

• There are a lot of minor mistakes, I would strongly suggest going through the work again to make the work more readable.