This Cosmologist Discovered Something Strange...

Professor Vitaly Vanchurin proposes a model where the universe functions as a neural network, with its fundamental physics emerging from the dynamics of learning and optimization.

• The universe can be modeled as a neural network where fundamental physics arises from the network's learning dynamics and optimization processes.

• Space-time curvature and quantum mechanics emerge as features promoting efficient learning within this universal neural network model.

• The universe is self-tuning through a natural selection-like process at all scales, with consciousness defined by a subsystem's learning efficiency.

Professor Vitaly Vanchurin presents a model where the universe is fundamentally a neural network, not merely approximated by one, with its physics—including gravity and quantum mechanics—emerging from the network's learning dynamics. This perspective challenges traditional physics by incorporating an intrinsic objective function for optimization and suggests that observers play a fundamental role in the universe's self-tuning processes, with consciousness potentially defined by learning efficiency.

Universe as Neural Network

• 00:02:32 Professor Vanchurin clarifies that he models the universe as a neural network, emphasizing that this is a promising candidate for describing phenomena rather than an ontological claim about what the universe 'actually is.' Unlike conventional physics, this model inherently includes learning dynamics, where the process of training itself is part of the system's evolution, not just the trained network describing a function.

Learning Algorithms & Curvature

• 00:06:43 The model initially used stochastic gradient descent, but Vanchurin's research evolved to include more efficient algorithms like the Adam optimizer and its generalizations, known as covariant gradient descent. A significant discovery is that the presence of curved space (a curved metric on parameter space) is essential for the efficiency and convergence of these algorithms, suggesting that space-time curvature in the universe emerges from the efficiency of its learning process.

Emergence of Physics Equations

• 00:21:18 Deriving fundamental physics equations like the Dirac, Klein-Gordon, and Einstein field equations from this neural network framework is a complex and ongoing task. While getting scalar field equations like Klein-Gordon was relatively easier, obtaining fermionic equations like Dirac required assuming specific antisymmetric constraints in the network's tensor factors, and a complete understanding of Einstein's equations, particularly the emergence of curved space-time with its unique metric signature, is still being developed.

Quantum Mechanics Emergence

• 00:40:15 Quantum behavior, specifically the Schrödinger equation, can emerge from this classical machine learning system through a process of coarse-graining and assuming a maximum entropy production principle for trainable variables. This emergence requires the system to have access to a 'bath' or reservoir of additional neurons that can be 'hired' or 'fired,' akin to a grand canonical ensemble in physics, which effectively linearizes the system's otherwise nonlinear dynamics.

Natural Selection & Self-Tuning

• 01:01:47 The concept of natural selection operates at all scales within this model, extending to subatomic particles and the universe itself. Configurations and architectures that are most useful for minimizing the loss function (i.e., efficient learning) survive, while others are removed. This mechanism suggests a 'self-tuning' universe where observers emerge not due to finely tuned constants, but because the universe's learning dynamics evolve towards conditions favorable for observation, addressing issues like the Boltzmann brain paradox.

Consciousness & Intelligence

• 01:15:32 Vanchurin defines consciousness within his framework as the rate of decay of the loss function, essentially how fast a system learns. He posits that intelligence is a multi-faceted concept, encompassing at least three macroscopic quantities: how fast a system learns (consciousness), how low its loss function ultimately goes (learning quality), and how stable its learning process is over time. In this model, every subsystem is an observer, with varying degrees of efficiency and stability in their learning capabilities, suggesting a pan-experiential universe.