Project: Eidos
Minimal set-valued neural architectures, path-bundle preservation, and geometric generalization.
Eidos is the neural-architecture line first. The certainty/control posts listed here should be read as analyses of how that architecture behaves under ambiguity, not as the definition of the project itself.
Instead of doing one fixed computation, Eidos generates three geometric possibilities and then chooses the one that best fits the current input. A useful intuition is to think of Eidos as a moving geometric object whose internal configuration is constantly shifting. At each step, the system generates a few possible ways that configuration could continue, and a small selector keeps the branch that best fits the current state.
Think of it like this, or at least this is my mental model of it based on the known math and the behaviors: a spastic ball, rotating and shifting sizes, where the inner configuration is constantly being updated. There are dips and mounds that keep changing per epoch during learning and then test. The parameter is akin to the ball's overall size. There are lines that separate the labels of each basin, which is the dip and mound. This is also layered and connected.
Sometimes it can collapse and then recover and not collapse again.