[title: line-fit] better vis through math and cs!
DimReader
- joint work with Rebecca Faust and David Glickenstein
- published at IEEE VIS 2018
[slide-data: backgroundImage /talks/boston-2018/images/tsne/cnn_embed_6k.jpg] [title: line-fit bg-black] t-SNE plots are beautiful, but what are they actually showing us?
Source: Andrej Karpathy
a simpler example: the iris dataset
| Sepal Length | Sepal Width | Petal Length | Petal Width | Species |
|---|---|---|---|---|
| 5.1 | 3.5 | 1.4 | 0.2 | setosa |
| 4.9 | 3.0 | 1.4 | 0.2 | setosa |
| 4.7 | 3.2 | 1.3 | 0.2 | setosa |
| ... | ||||
| 7.0 | 3.2 | 4.7 | 1.4 | versicolor |
| 6.4 | 3.2 | 4.5 | 1.5 | versicolor |
| 6.9 | 3.1 | 4.9 | 1.5 | versicolor |
| ... | ||||
| 6.3 | 3.3 | 6.0 | 2.5 | virginica |
| 5.8 | 2.7 | 5.1 | 1.9 | virginica |
| 7.1 | 3.0 | 5.9 | 2.1 | virginica |
| ... |
[title: empty]
[title: empty] [attribute: id slide-iris-tsne]
t-SNE visualization
how?
- we use automatic differentiation to compute how the projection would change if the data changed
- the rest is linear algebra and classic computer graphics algorithms
DNNs through the Grand Tour
- under revisions for distill.pub
- joint work with Mingwei Li and Zhenge Zhao
idea: DNNs are almost linear
-
every layer is a linear map followed by a simple elementwise operation
-
linear maps “just” change the basis under which the dnn works
-
pick a DR algorithm robust to linear maps to see the neural network change the input layer-by-layer
example: grand tour of iris
- now do this for inputs as they go through a dnn
demo
takeaways
you can get through a lot of slides by speaking fasti spend too much time hacking javascript slides- you can do better vis with better math
- thanks