Singular Learning Theory
2024-10-29 — 2026-11-16
In Which Watanabe’s Singular Loss Surfaces Are Navigated, Algebraic Geometry Is Invoked to Tame Uninterpretable Parameters, and Local Learning Coefficients Are Estimated to Replace the Intractable RLCT.
On this page
Placeholder.
As far as I can tell, a first-order approximation to (the bits I vaguely understand of) Singular Learning Theory is something like:
Classical Bayesian statistics has a good theory of well-posed models with a small number of interpretable parameters. Singular Learning Theory is a theory of ill-posed models with a large number of uninterpretable parameters, which provides us with a model of Bayesian statistics by using results from algebraic geometry about singularities in the loss surface.
Why might we care about this? For the moment I am taking it on faith. Since attending ILIAD2 I am relatively more optimistic about the potential for this program of research to go somewhere.
Jesse Hoogland, Neural networks generalize because of this one weird trick:
Statistical learning theory is lying to you: “overparametrized” models actually aren’t overparametrized, and generalisation is not just a question of broad basins.
1 Local Learning Coefficient
Resources recommended to me by Rohan Hitchcock:
- The upshot is that Jesse Hoogland and Stan van Wingerden argue that we should care about model complexity, and the local learning coefficient (Lau et al. 2025) is arguably the correct measure.
- The global version is the RLCT, but as far as I can tell, it is intractable.
I probably have some alpha in estimating this value, so I am doing so — see Estimating the local learning coefficient of neural networks.
2 Fractal loss landscapes
Notionally, there might be a connection to fractal loss landscapes. See Fractal dimension of loss landscapes and self similar behaviour in neural networks.
3 Use in developmental interpretability
I don’t understand this step of the argument, but see Developmental Interpretability by Jesse Hoogland and Stan van Wingerden, and maybe read Lehalleur et al. (2025).
4 Questions I would like to know how to answer
- Can we use LLC as an optimal design objective, crafting desired optima directly in loss space by altering loss functions and/or architectures dynamically?
- Can we filter to estimate the LLC online? It looks a lot like sparse Kalman filtering — just sayin’.
- The Bayesian formalism. How does the notional “Bayesian” update work in practical neural networks that are not trained by posterior updates? Does it matter? Can we use standard Bayes-by-backprop arguments? What happens when our training process really pushes the analogy, e.g., when we have synthetic data generation in a neural distillation?
- How biased is LLC estimation for non-trivial networks, i.e., non-linear ones? I ran some sims and made the LLC estimates of SGHMC and SGLD diverge substantially. Is that worrisome?
- How would belief propagation work in a singular setting? What even is the natural space of priors generating local landscape geometries?
5 Incoming
From SLT to AIT: NN generalisation out-of-distribution
This post derives an upper bound on the prediction error of Bayesian learning on neural networks. Unlike the bound from vanilla Singular Learning Theory (SLT), this bound also holds for out-of-distribution generalization, not just for in-distribution generalization. Along the way, it shows some connections between SLT and Algorithmic Information Theory (AIT).
The Developmental Interpretability site is mostly about SLT, and it runs a thriving Discord server on the topic.
Alexander Gietelink Oldenziel — Singular Learning Theory
metauni’s Singular Learning Theory seminar
-
Timaeus is an AI safety research organisation working on applications of Singular Learning Theory (SLT) to alignment.
Brennan Dury has an interesting agenda and explanatory style:
Concretely, index a sequence of data sets by \(t\). For each loss \(j\), define the empirical loss
\[ K_{j,t}(w) = \frac{1}{n_{j,t}} \sum_{i=1}^{n_{j,t}} f_j(X_{j,i}, w). \]
Here \(n_{j,t}\) is the sample size for loss \(j\), and the observations \(X_{j,1}, X_{j,2}, \ldots\) are independent with distribution \(\nu\) within each stream. Consider the family of Gibbs distributions
\[ \Pi_t(dw) = \frac{1}{Z_t} \exp\left[ -\sum_{j=1}^{\ell} n_{j,t} \beta_{j,t} K_{j,t}(w) \right] \varphi(w)\, dw, \]
where the weights \(n_{j,t}\beta_{j,t}\) can grow at different rates, and \(\beta_{j,t}\) is the inverse temperature for loss \(j\). The normalizing constant \(Z_t\) and its derivatives contain important information about the system. With one loss, its leading decay is governed by the RLCT. With several losses, the corresponding exponent depends on their relative weighting.
