Visualizing Data Using t-SNE

Google Tech Talk
June 24, 2013
(more info below)
Presented by Laurens van der Maaten, Delft University of Technology, The Netherlands

ABSTRACT

Visualization techniques are essential tools for every data scientist. Unfortunately, the majority of visualization techniques can only be used to inspect a limited number of variables of interest simultaneously. As a result, these techniques are not suitable for big data that is very high-dimensional.

An effective way to visualize high-dimensional data is to represent each data object by a two-dimensional point in such a way that similar objects are represented by nearby points, and that dissimilar objects are represented by distant points. The resulting two-dimensional points can be visualized in a scatter plot. This leads to a map of the data that reveals the underlying structure of the objects, such as the presence of clusters.

We present a new technique to embed high-dimensional objects in a two-dimensional map, called t-Distributed Stochastic Neighbor Embedding (t-SNE), that produces substantially better results than alternative techniques. We demonstrate the value of t-SNE in domains such as computer vision and bioinformatics. In addition, we show how to scale up t-SNE to big data sets with millions of objects, and we present an approach to visualize objects of which the similarities are non-metric (such as semantic similarities).

This talk describes joint work with Geoffrey Hinton.

[WayneStidolph]

Nicely presentation - I'm a naive layman, but I was able to follow along and see how this is a useful technique. Thank you for sharing!

[chriscanal]

That visualization at 20:31 is so baller. Such savage domination over the competing algorithms

[lebesgue-integral]

One of the best presentations I have ever seen in ML

[RalphDratman]

Wonderful talk, very clear, giving by a wide margin the greatest real-world impact of any Google talk I have seen.

[casemoy]

Thank you so much. One of the best talks I ever listened to!

[juliankuo]

Such an impressive work which i should carefully read before!

[peterfranken]

Very clear and insightful presentation. I cant wait trying it out myself.

[dennishain]

Great stuff - I'm thinking it's time to get more deep into t-SNE for more insights about our data.

[xintongbian]

didn't get hinton's introductory talk, what was the four and the 12 that he was talking about?

[yingbeesweden]

hello my friend nice film and like, nice to meet you

[jieqiangwei]

great performance with simple ideas!! Fantastic!

[Nathanielmhld]

Was that first question asked by UC Berkeley's Jon Deniro????

[jteoh]

I am interested in looking at the interactive 3D tool on your website visualizing your data, do you have a direct link to the interactive plot that you can share? Thanks in advance

[inferno-jmrd]

can someone explain how t dist separates dissimilar points to be modelled far @20:10 ?

[alanwang]

That's why Google is the best company on earth

[phsamuelwork]

I assumed that the quadtree (27:06) is built for the original point set x_i in the high dimensional space. Can anyone explain how this can be done for points lied beyond 2D?

[gaaligadu]

Hi Guys, can someone please explain why symmetric probability is Pij = (Pi|j + Pj|i)/2N and not Pij = (Pi|j + Pj|i)/2 ?

[nikhilsrajan]

*HOLD TIGHT t-SNE*
He's got a pumpy.
(big ting)

[DavidAKZ]

What is a "high dimensional" object ?

[sayajujur]

why everyone using deep learning for image or text?
I want to use deep learning (and use t-SNE for visualization) on bioinformatic dataset I've collected. that dataset is, I can say larger version of IRIS dataset with 512*16 . How to do classification show it in t_SNE?