Artist Statement
The t-SNE Motion Sculpture is meant to be the visualization of the paths that different embeddings of words take when their dimensionality is reduced, in this case, from 100 to 3 dimensions. This process uses a machine learning algorithm called t-SNE.
See the term explanations at the bottom if you need to
Aside from the gifs and images in this repository the project is actually the app itself, which shows customization and exploration far beyond the wiki and the README. The app itself, when run features a gui, that allows the user to specify the parameters of the algorithm and run it again on the same data with new parameters. This allows the user to explore the different combinations of parameters and how it affects the output
There are also a few different modes of viewing the visualization. There are options in the GUI that make the last 10 points in each path connect to one another in random order, and there is another option that allows for hiding the path as well. Having everything turned off allows the user to see only the final points in the 3D space.
Connections and paths:
Just connections:
Just paths:
Just points:
The colors for the piece are somewhat arbitrary. I initially wanted it to just be functional so I made the colors for all of the different points be random so that you could see a path and find the corresponding point. I thought that I would figure out the colors after I perfected the actual form of the piece. See the Future Additions section.
One thing that was a little out of my control is the fact that the t-SNE algorithm is implemented to initialize itself to a random seed. So every time the algorithm is run, it results in a (slightly) different shape. I would be able to control this by having the points preprocessed.
The three view options were mostly as a prototyping method so that I could experiment with different looks for the program.
The paths were the original option. It wouldn't be a motion sculpture without them. I chose to make them so thin because when they were thicker the result looked a little less clean and more like someone was drawing with crayons in 3D space. The paths themselves are formed from curves rather than straight lines so that they aren't jagged and jarring. I wanted the forms to be smooth like in the Tensorflow embedding projector project.
The connections option was to show what kind of volume the the points would take up it they'd been turned into an actual mesh. However It does cloud a lot of the form itself because the white is just white and doesn't have a real material that could give the depth of the lines. One thing that I'm open to exploring is giving those connections volume and seeing how they would interact as a material with an artificial light source.
Word Embedding - This is a numerical representation (usually a list of decimal numbers, a vector) of what a word means in the context from which it is taken. In this case, these word embeddings are based of the wikipedia corpus. The algorithm that turns an input body of text (corpus) into this collection of embeddings for each word in that corpus, is called Word2Vec. The embeddings of different words can then be compared with distance (or other metrics) computations to find how close their meanings are. For example: if two word vectors are close to each other then it usually signifies that their meanings are also close. The purpose of this algorithm is to be able to group words together based on their meanings inside of a vectorspace.
Dimensionality - Dimensionality in this case is the number of numbers that represent a word in an embedding. In this case the original dimensionality is 100. Then the t-SNE algorithm brings it down to 3 so that we can visualize the clusters of words in a space that we can understand.
Motion Sculpture - This piece was meant to be a parametric art piece that was supposed to generate a new sculpture each time it was run, however due to the nature and complexity of it I was unable to make it be as customizable as I wanted. Thought the general idea of a motion sculpture can be seen here. It is the embodiment of a path that an object takes through space. This piece shows that path, but it lacks the solid, sturdy quality of a true sculpture. Ultimately I would like to allow for the creation of solid, less eccentric forms from this process.
This project was meant to be a culmination of the knowledge that I've learned over the semester with respect to machine learning and art. I explored such topics as perceptrons, Neural Networks, Convolution, Image processing, and finally, before this project, dimensionality reduction (t-SNE).