Data, the Ballerina
The world is complex – the features of a situation escape human reason. To best understand, we must use our faculties with the strongest heuristics: our capacity to interpret other humans. There are many examples of doing this within the arts – such as dance, which projects complex human interactions onto body motions. To understand data, we should make it dance.
Interpretation of complex events is a human weakness. We have limited capacity to understand causation in the modern world. This issue is compounded by our two main techniques of addressing this complexity:
- Numerical representations are accurate, but hard to interpret qualitatively.
- Traditional visualization techniques rely on our senses of geometry and color, occasionally enriched with iconography. This is inherently low-dimensional.
However, both fail to utilize our greatest capabilities – the interpretation of human actions. Humans show a surprisingly rich capacity to interpret the actions of other humans. This is driven not only be an emphasis on social learning, but by specialized regions within our brains. By stepping away from the explicit world of numbers or the rigid world of geometry, we can train our intuition to interpret events holistically.
The world is coplex in ways that don’t align to human intuition. So we must make our data analysis more interpretable by aligning data with human capability – eg, by teaching data to dance. The long term vision is ambitious, but simple: to generate a ballet which represents Wall St behavior over the course of a year. To lean into our habit of anthropomorphizing in a rigorous way and show the interplay between stocks as an interplay between (rendered) ballerinas.
There is substantial work in creating art via machine learning. Deep neural networks can be used to encode complex data (such as art) as points on a manifold or to generate facimiles given a point on the manifold. By training multiple networks together, such as in a cycle GAN model, translations between domains can be achieved. The proposal here to teach data to “dance” can be thought of as finding the latent structures within financial data and using those to generate a performance – to map the rises and falls, the timing of movements to features of financial data.
Currently being explored is the application of effective semantics to this process – by extracting the “internal language” of both finance and dance, then aligning the translation between them based on that structure. This technique more closely replicates the human approach to matching “expressions” within each domain.