The key to writing a survey paper is to recognize that the reader
already has access to the original papers that are being surveyed.
So, if the survey simply has a section summarizing each of the original
papers, then the survey is not really adding value.  The reader could
gain the same benefit by reading the abstract, introduction, and
topic sentences of each of the original papers.
    Similarly, if you are tempted to explain some of the details
in the original paper, then just refer the reader to directly
reading, for example, "Paper 2, Section 3" for detailed.
One of the tests of a good survey paper is whether you can condense
the material into a 5-minute oral presentation.  If you can't condense
it, then maybe you're trying too hard to present three distinct
papers, and 5 minutes is not enough to present three unrelated papers.
This is then a clue that you're on the wrong track.
    The creativity in a survey paper is searching for a common
theme (perhaps by also reading a Wikipedia article or blog for
background), and then commenting on how each paper addresses
the common theme.  This is the part that can be fun, once you
realize the target that the reader of the survey is looking for.
Otherwise, "why does the reader need you, anyway?", to describe the
point in a more humorous vein.

Researchers read the literature in a top-down or hierarchical manner.
First, they read the title, to decide if they should look more closely.
The title should have the right keywords, but usually nothing more.

If they like the title, they will read the abstract.  In the case of
a survey, the abstract should have the key idea, and a sentence or two
about why it is important.  Normally, the abstract fits in a single
paragraph.

If they like the abstract, they will reach the intro.  In the intro,
if you read the topic sentences, then it should tell a story.
A typical story (e.g., for deep learning) might have the following
topic sentences:

1. Deep learning allows more accurate learning by having more levels
   in the deep learning network.

2. Deep learning has become practical through the use of GPUs and
   other hardware accelerators.

3. A key problem in deep learning is how to represent a particular
   application within a deep learning network.

4. Two ways in which applications may be mapped to a deep learning are
   through mapping a feature space, or through mapping a two-dimensional
   set of pixels.

5. The performance of deep learning varies drastically according to
   the representation of the application.  Higher performance tends
   to be exhibited when ....

And now that you have your story based on topic sentences in the intro,
you can use the body of the paper to provide more details of that story.

Finally, a key question that comes up is how to write one survey
paper that integrates three widely varying papers.  Conceptually,
you will have read each paper in depth before going on to the next one:

PAPER:
  1   2    3
  _   _    _
  |   |    |
  |   |    |
  V   V    V

Now, your own survey paper should concentrate on breadth:
  1   2    3
  |--------->
  |--------->
  |--------->
  |--------->

For example:
1.  Deep learning uses more levels.  The number of levels can vary from
    X to Y, depending on the application.
2.  Deep learning typically uses GPUs for hardware acceleration.  But at
    least one Deep Learning approach uses the Intel Xeon Phi many-core CPU.
3.  The problem of mapping an application to a deep learning network
    is the key to being able to solve that application.  Some examples
    of applications that can map easily to a deep learning network are ....
4.  Tyipcal application representations include feature spaces and
    pixels in Euclidean space.  We see that the representation in a
    deep learning network varies radically for these two cases.
5.  The performance of deep learning also varies radically for
    different applications.  While this field is still evolving,
    current experiments have shown that feature space approaches
    tend to exhibit higher performance than geometric Euclidean spaces,
    when comparing deep learning approaches to standard approaches in
    the corresponding application domain. [OR WHATEVER YOU CONCLUDE INSTEAD.]