Compliments of a discovery by Futurism, the paper The Autodidactic Universe by a smorgasbord of contemporary science and technology thinkers caught my attention for several reasons. First was Jaron Lanier as a co-author. I knew Jaron’s dad, Ellery, when I was a researcher at NMSU’s now defunct Computing Research Laboratory. Ellery had returned to school to get his psychology PhD during retirement. In an odd coincidence, my brother had also rented a trailer next to the geodesic dome Jaron helped design and Ellery lived after my brother became emancipated in his teens. Ellery may have been his landlord, but I am not certain of that.
The paper is an odd piece of kit that I read over two days in fits and spurts with intervening power lifting interludes (I recently maxed out my Bowflex and am considering next steps!). It initially has the feel of physicists trying to reach into machine learning as if the domain specialists clearly missed something that the hardcore physical scientists have known all along. But that concern dissipated fairly quickly and the paper settled into showing isomorphisms between various physical theories and the state evolution of neural networks. OK, no big deal. Perhaps they were taken by the realization that the mathematics of tensors was a useful way to describe network matrices and gradient descent learning. They then riffed on that and looked at the broader similarities between the temporal evolution of learning and quantum field theory, approaches to quantum gravity, and cosmological ideas.
The paper, being a smorgasbord, then investigates the time evolution of graphs using a lens of graph theory. The core realization, as I gleaned it, is that there are more complex graphs (visually as well as based on the diversity of connectivity within the graph) and pointlessly uniform or empty ones. Fair enough, similar to some of my own work on evolving neural networks and examining the distribution of connectivity weights that allows for solution-solving behavior. I thought it was interesting in 1994 but it had no practical impact, nor did I pursue it further. There are some ideas for extending the graph theoretic approach in the paper that may lead to some new research, however.
The paper then moves into discussions of biological evolution as a learning model and an analysis of how previous states of the systems influence future changes. Like I said, it’s a smorgasbord of people and ideas. Or maybe more of a buffet where each co-author brought along a dish based on the main theme of learning. There are tasty options for everyone and most interests.
The initially stated main goal is, however, to discuss whether the universe itself is autodidactic, a self-learner of all the laws and constants that we now observe. This is finally pulled together in the conclusion but there are hints about the evolution of this idea along the way in the discussion of cosmological ideas that involve metaphors of Darwinian evolution or multiverse-style ideas, as well as reversibility and the time evolution of systems.
The end result suggests there may be impact on quantum simulators or on the construction of special types of quantum computers. Fair enough. The paper is a long journey that could easily have been broken up into more focused chunks and presented individually, but it makes for an easy scan watching only for the relevant differentiators and connective tissue that binds the sections together.