Vanchurin's Neural Network Universe
Physics is what a learning neural network produces at large scales. The substrate is not fields or strings but a network of nodes and connections that learn over time, and what we call quantum mechanics and gravity are emergent descriptions of its dynamics.
Placeholder for a 3D visualisation of Beneath the Standard Model. The interactive scene will land in Phase 3. Most approaches to a final theory try to quantize gravity or unify forces at high energy. This family takes a different tack: physics as we know it emerges from something else underneath. The substrate might be a learning neural network whose dynamics produce quantum behavior, or a fundamentally classical spacetime that requires modifying quantum mechanics instead. The proposals disagree sharply on what the substrate is, but they share the conviction that the standard pair of quantum theory and general relativity is not the foundation.
In one sentence
The universe at its deepest level is a vast neural network whose learning dynamics give rise to both quantum mechanics and gravity.
The claim
Vanchurin's hypothesis flips the usual relationship between neural networks and physics. Instead of using neural networks as models of physical systems, he proposes that physical systems are themselves instances of a cosmic neural network. The fundamental substrate is not fields, particles, or strings, but a network of nodes and connections that learn over time.
The patterns this network produces at large scales look, from our perspective, like particles, forces, and spacetime. Vanchurin argues that the learning dynamics of a generic neural network produce equations resembling quantum theory in one limit and classical or relativistic dynamics in another. If something like this is right, then learning and information processing are as fundamental to physics as mass or charge.
The family stance
Familiar physics is not fundamental but emerges from a deeper substrate. Different proposals identify that substrate differently, but they share the claim that quantum mechanics and gravity are not the bottom layer of reality.
Predictions
- Any physical phenomenon should be reproducible by a suitably constructed neural network. Finding a robust counter-example would falsify the hypothesis.
- At certain scales, the effective physics should show subtle, scale-dependent departures from standard quantum mechanics or general relativity, traceable to limits of the neural network approximation.
- The framework implies specific non-local hidden-variable structure underlying quantum phenomena. Future experiments that constrain such structure would directly test the proposal.
Evidence
- Vanchurin shows that, under certain assumptions, equations governing the network's trainable parameters during learning can approximate equations resembling quantum mechanics in one regime and classical dynamics in another.
- Neural networks are known to be universal approximators, capable of reproducing complex dynamical behaviors, which motivates the idea that they could approximate known physical laws.
- A growing body of work in machine learning shows neural networks efficiently emulating quantum many-body and gravitational systems, supporting the idea that neural-network-like structures can encode the same patterns as standard theories.
Counterpoints
- The hypothesis has few concrete quantitative predictions that differ from standard physics, so most physicists view it as a philosophical proposal rather than a tested theory.
- Saying everything can be modeled as a neural network risks being too broad to falsify, because neural networks are flexible enough to approximate many patterns.
- The claim that quantum mechanics and general relativity emerge from learning dynamics has not been derived in a way that reproduces all the detailed structure (symmetries, spectra, renormalization) of known theories.
- The framework replaces one mystery (what is the fundamental ontology?) with another (why this particular network architecture and learning rule?), without obvious guidance from observation.
- Compared with developed quantum gravity programs, the neural network picture has very little engagement in the peer-reviewed physics community, so it remains a fringe idea awaiting scrutiny.
Variants in this family
▸Go deeperTechnical detail with proper terminology
Vanchurin splits the neural network's degrees of freedom into trainable variables (weight matrices, bias vectors) and hidden variables (neuron state vectors), then studies their stochastic evolution. Near equilibrium, the effective dynamics of trainable variables approximate Madelung-form quantum equations. Far from equilibrium, they approximate Hamilton-Jacobi equations with free energy playing the role of phase or action.
He interprets the hidden variables as non-local hidden degrees of freedom reminiscent of Bohmian mechanics. Quantum states and their evolution are emergent descriptions of underlying network states, explicitly aligning the proposal with emergent or Bohmian quantum mechanics rather than Copenhagen interpretation.
The framework is cast in terms of a learning equation, where stochastic gradient-like evolution in parameter space gives rise in different limits to effective classical and quantum field theories. Vanchurin speculates that the Einstein-Hilbert action appears as an effective description of the free-energy production associated with learning.
The proposal is a cousin of 't Hooft's cellular automaton interpretation, where a deterministic discrete substrate produces quantum behavior at larger scales. The substrate here is a stochastic neural network rather than a cellular automaton, but both treat quantum mechanics as emergent from underlying variables with non-local structure.
References
Spotted an error? Have a source to add?
Prefer email?
You can also send a prefilled email with the variant URL already filled in.