So in summary we've covered
a lot of ground in this module. So we've returned to this task of
retrieving a document of interest. And have formulated this in terms of
performing nearest neighbor search. We reviewed our nearest
neighbor search algorithms for performing one nearest neighbors search
as well as K nearest neighbor search. But we got into some of
the nitty-gritty of these algorithms. In particular,
thinking about how we represent our data. So we focused here on different
representations of documents. But we also talked about the really,
really critical idea of, how do we compute distances
between two different data points? And we went through a set of different
ways to compute these types of distances and discussed
the trade-offs that we're faced with. And the fact that there's
really no one answer for what's the right data representation and
what's the right distance to use. And that's where a lot of
domain knowledge comes in. It can have a lot of impact on
the performance of the algorithms. And you really have to think
very critically there. But then, for the rest of the module
we talked about ways to think about performing really efficient in nearest
neighbour search to help handle the fact that the brute force search. When you have lots and lots of
observations, can be really, really, really computationally intensive. So we talked about KD-trees. That's a very efficient data
structure that we can use for performing our nearest neighbor search. And we showed when it can be useful, but we also said that it's
struggled in high dimensions. So then we turned to this idea
of locality sensitive hashing, which only provides
approximate nearest neighbors. But we motivated that that can actually
be reasonable in a really large set of different applications. But, this locality sense of
hashing is really, really, really simple to implement. And can be very efficient. And provide good approximations that
have probabilistic guarantees on their approximations made. So, coming out of this module, you really have some real powerful tools
that are useful in just a huge number of retrieval tasks going
beyond document retrieval. Lots and lots of applications
where you have some query and you want to find some similar object. And all the tools we've discussed in
this module generalize to these other applications as well. And so, I'll just leave you with a summary
of a set of things that you should be able to do now that you've seen this module. [MUSIC]