Agenda
A central question of out-of-distribution generalization is as follows. Suppose the set of parameters that achieve minimal training loss is not a singleton, and different optimal parameters implement different behavior outside the training distribution. Which solutions are favored by the training algorithm?
Singular learning theory answers, at least when the training algorithm is Bayesian inference and certain regularity conditions hold, that solutions with minimal real log canonical threshold (RLCT) are asymptotically favored. You can think of the RLCT as a measure of local geometric degeneracy of the loss function. Researchers in many fields have proposed notions of “effective dimensionality” to capture the intuition that not all parameters matter equally, so learning good points in parameter space may be easier than the literal dimensionality of parameter space suggests. It may help to think of RLCT as the notion of effective dimensionality that appears in the asymptotic prediction error of Bayesian inference, , where is the minimal RLCT and is the number of samples.
Now suppose two population-loss minimizers have RLCTs . For small neighborhoods containing , the Bayesian posterior assigns times more mass in favor of , up to logarithmic factors. This turns the central question above into an algebraic geometry question: for a given behavior out of distribution, is the minimal RLCT among solutions which exhibit that behavior smaller or greater than the minimal RLCT among solutions which do not exhibit that behavior?
Suppose you train your AI not to cheat its tasks. The AI could learn the policy “do not cheat” or learn the policy “do not cheat if and only if I am being trained.” Both policies have the same behavior on the training distribution, so loss minimization alone cannot predict which policy the AI learns. In the Bayesian setting, singular learning theory predicts an asymptotic preference for either policy if it has smaller RLCT than the other policy. If we could understand which policies minimize RLCT, we could better predict which behaviors Bayesian inference favors. And if we could design architectures and loss functions to affect those RLCTs, we could shape that preference.
My current agenda seeks to understand how designing additional loss functions can affect the geometry of the loss landscape. Concretely, index a sequence of data sets by . For each loss , define the empirical loss
Here is the sample size for loss , and the observations are independent with distribution within each stream. Consider the family of Gibbs distributions
where the weights can grow at different rates, and is the inverse temperature for loss . The normalizing constant 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.
My framework describes this dependence using polyhedral geometry. For weights expressed in the form , the leading exponent is piecewise linear in the vector . This gives a precise sense in which the system has a phase diagram: the limiting normalized free energy can become nondifferentiable as the relative growth rates change.
The derivatives of with respect to the weights give the posterior expectations of the individual losses. The framework therefore describes how strengthening one objective affects both its own expected loss and those of the other objectives.
So far, my main application is to prove that the RLCT estimator proposed by Lau et al. in The Local Learning Coefficient: A Singularity-Aware Complexity Measure is consistent. Watanabe’s work in A Widely Applicable Bayesian Information Criterion establishes the consistency of a closely related estimator assuming the prior is compact. For computational reasons, Lau et al. instead use a Gaussian localization factor
to softly concentrate the prior near a point .
There are two competing requirements. The quadratic penalty must become strong enough to concentrate the posterior near . But if its weight grows too quickly, it can dominate the posterior, causing the estimator to average the original loss under a distribution shaped primarily by the localization penalty rather than by the original loss.
My result identifies a broad range of growth rates satisfying both requirements. Write for the primary loss’s weight, where is the inverse temperature. Under the theorem’s other regularity conditions, it is sufficient that
The third condition allows the result to not assume that the model satisfies the relatively finite variance assumption. The first two conditions permit the localization weight to grow faster than logarithmically with respect to while remaining slower than every positive power of .
More generally, my result gives conditions under which finitely many auxiliary losses make the posterior distribution of the scaled primary loss asymptotically Gamma:
where the Gamma distribution has shape and rate , and is the minimum RLCT of the primary population loss along the common zero set of all the population losses. Here convergence means weak convergence of the conditional posterior law, in probability over the training data.
This application illustrates one way auxiliary losses can shape Bayesian selection: they can restrict where the posterior concentrates while preserving the primary loss’s asymptotics locally. Further work in this agenda may use the phase diagram to understand regimes where multiple loss functions are asymptotically influential. For instance, splitting one loss function into a mean term and an error term, we may better understand Bayesian asymptotics when the mean term and error term have very different geometry. We may also better understand when training on one dataset aids fine tuning on another dataset.