The "dream" has historically been conceived in the West as internal, entirely private, impossibly personal. This is at odds with much of the world, where dreamers and dream workers see the dream either as a parallel and continuously existing landscape (i.e. Aboriginal Australian Pintupi) or see the dream as internal spaces which must be aligned and made continuous for the creation of a cohesive society (i.e. Ongees of the Andaman Islands).
Dream data in the US is typically used to separate us, as another kind of data for categorizing and capturing; this is what people dream when they are young, or old, or blind, or Republican. There is great interest in this sort of data, which has shown sensitivity to things which subjects might otherwise want kept secret, by powers like the British Colonial Office centuries ago or Bell Labs just this year.
This project is interested instead in the creation of a continuous dream. Of dream walking. It is an automated, algorithmic negotiation between stories, a shuffle so that one end meets another's beginning. Given 1000 dreams, it will reorder them such that they create one continuous dream. Given my dream and yours, whoever you may be, it will create a path that walks, hops lily pad to lily pad, dream to dream, from one story to another until I arrive in your dreamscape.
Interpolating between dreams in such a manner requires establishing meaningful relationships between a multitude of dreams. We utilize the BERT framework, a powerful transformer with an extensive understanding of natural language, to align thousands of dreams according to common topics. This topical clustering happens beyond simple pattern recognition. BERT’s deep learning approach to natural language processing makes it capable of understanding complex semantics. The result is a high-dimensional cloud of dreams where spatial proximity between data points translates to semantic relatedness. Dreams with comparable content are packed together, forming small and big communities of common themes. Moving through its axis allows us to explore these communities.
In order to make this cloud easily accessible to humans, its dimensions are reduced to two or three via Universal Manifold Approximation and Projection, or UMAP, before a final clustering process through HDBSCAN. The resulting point cloud can be almost treated as a geographic entity. To interlink two dreams, one can simply walk a path in-between and inspect the dreams situated along the way. To do so, we developed a path-finding algorithm designed for a balance between locally and globally optimal paths. After selecting a desired start and end dream, the system yields a list of real dreams which semantically approach each other. Following this path of dreams does not only offer us a better understanding of the cognitive and emotional distance between two dreamers but also a potential guide for connecting their mental spaces. Where we once seem impossibly distant, each in our dreamt spaces a world apart, here we see a path composed of small, familiar steps, connecting us.