String Theory from Infinite Width Neural Networks

The Big Picture

Imagine training a neural network so wide (so many neurons per layer) that it stops behaving like a neural network and starts behaving like a physical field, the mathematical object physicists use to describe forces and particles. This isn’t a metaphor. It’s a theorem, and researchers at Northeastern University are using it to do something that sounds impossible: build string theory out of neural networks.

String theory proposes that the fundamental constituents of reality are tiny vibrating strings rather than point particles. It remains one of the few candidates for reconciling quantum mechanics with gravity, but computing how strings interact requires elaborate mathematical machinery that physicists have spent decades developing. Samuel Frank and James Halverson have now shown you can reproduce those results by computing the statistics of carefully designed neural networks, deriving the Veneziano and Virasoro-Shapiro amplitudes, the formulas predicting how strings scatter, directly from network statistics.

Key Insight: The variance of a neural network’s output weights directly controls string tension. Tune one hyperparameter in your network, and you tune the fundamental physical scale of the strings.

How It Works

The bridge rests on a deep result from machine learning theory. Send a neural network’s width to infinity and its outputs become draws from a Gaussian process, a random function fully characterized by a probability distribution. This is the Neural Network Gaussian Process (NNGP) correspondence, and it means an infinite-width network’s statistics are completely described by a free field theory. Frank and Halverson build on the Neural Network Field Theory (NN-FT) program, which exploits this correspondence not for machine learning, but for physics.

Bosonic string theory, after fixing coordinates on the string’s surface, has two ingredients: the free boson (describing string motion through spacetime) and the bc-ghost system (additional fields required to keep the equations consistent once that coordinate choice is made). The researchers built explicit architectures for both:

• Free boson architecture: Each spacetime coordinate $X^\mu(z)$ is a sum over $N$ neurons, each contributing a cosine of a complex frequency $W_i$ and random phase $c_i$, weighted by Gaussian output weights $a^\mu_i$. Frequencies are drawn uniformly from an annular region in the complex plane, bounded by IR cutoff $\epsilon$ and UV cutoff $\Lambda$.
• Ghost architecture: The $b$ and $c$ ghost fields are represented by Grassmann-valued networks, networks whose weights obey anticommuting arithmetic (the mathematical structure required for fermions), using analogous random-phase constructions.

Averaging over the free boson network’s random parameters recovers the standard string propagator exactly: a logarithm of the separation $|z - w|$ in position space, a $1/p^2$ scaling in momentum space. The only non-regulator parameter is $\sigma_a$, the output weight standard deviation. This single number fixes the Regge slope $\alpha’ = \sigma_a^2 / (\Lambda^2 - \epsilon^2)$ and therefore the string tension $T \sim 1/\alpha’$. One dial stretches or compresses the strings.

The ghost system follows the same logic with Grassmann variables replacing real numbers. Averaging over those fermionic weights recovers the $\langle bc \rangle$ propagator, the $1/(z-w)$ pole every string theorist recognizes, up to corrections that vanish as the UV cutoff grows.

With both components built, string scattering becomes a computation in network correlators. Vertex operators encode the quantum state of each incoming and outgoing string as insertions in the network’s output.

The four-point tachyon amplitude, integrated over the worldsheet (the two-dimensional surface a string sweeps through spacetime), becomes an integral over the network’s parameter space. That integral produces the Veneziano amplitude: a ratio of gamma functions with characteristic crossing symmetry and Regge behavior, where high-energy scattering is exponentially soft rather than power-law divergent. The closed-string version yields the Virasoro-Shapiro amplitude. Both fall out cleanly from the network formalism.

Why It Matters

This result pulls in two directions at once. For string theory, it’s a new computational angle. Standard derivations live in the language of conformal field theory and path integrals. Frank and Halverson’s derivation lives in the language of random functions and covariance kernels, a different toolkit that may illuminate corners of the theory the standard approach obscures. That $\alpha’$ is controlled by a single variance parameter makes the physical interpretation unusually transparent: string theory’s fundamental length scale is literally the spread of a weight distribution.

For the NN-FT program, this is a landmark. Previous work had constructed free scalars, free fermions, supersymmetric theories, and conformal field theories in this framework. String theory, a theory of quantum gravity, is a big step up in ambition and complexity.

The result raises immediate questions. If bosonic string theory lives here, where do superstring theories live? What does finite-width physics look like, and does it correspond to string interactions? The researchers note their work is largely complementary to independent work by Robinson on neural network realizations of Virasoro symmetry, suggesting the field is converging on something real.

Bottom Line: A neural network’s weight variance controls string tension. String theory’s foundational scattering amplitudes can be computed as the statistics of a well-chosen random function, giving physicists a new computational and conceptual route into quantum gravity.