Lauren A. Hannah

Soheil Zibakhsh, Mohammad Samragh, Kumari Nishu et al.

The generation quality of large language models (LLMs) is often improved by utilizing inference-time sequence-level scaling methods (e.g., Chain-of-Thought). We introduce hyper-parallel scaling, a complementary framework that improves prediction quality at the token level. Hyper-parallel scaling computes and aggregates multiple output proposals for a single token from the model. We implement this concept in Mixture-of-Experts (MoE) models, which we refer to as Roster of Experts (RoE). RoE is a training-free inference algorithm that turns a single MoE into a dynamic ensemble of MoEs. RoE injects controlled stochasticity into the expert routing mechanism, enabling it to sample multiple diverse experts for each token and aggregate their outputs for a more accurate final prediction. To overcome the computational cost, we introduce an efficient batching strategy and a specialized KV-caching mechanism that minimizes compute and memory overhead. For example, RoE enables a 7B MoE model to match the performance of a 10.5B MoE model while using 30% less compute for inference. These gains are achieved without any fine-tuning of model parameters.

Louis-Francois Arsenault, Richard Neuberg, Lauren A. Hannah et al.

We present a machine learning approach to the inversion of Fredholm integrals of the first kind. The approach provides a natural regularization in cases where the inverse of the Fredholm kernel is ill-conditioned. It also provides an efficient and stable treatment of constraints. The key observation is that the stability of the forward problem permits the construction of a large database of outputs for physically meaningful inputs. We apply machine learning to this database to generate a regression function of controlled complexity, which returns approximate solutions for previously unseen inputs; the approximate solutions are then projected onto the subspace of functions satisfying relevant constraints. We also derive and present uncertainty estimates. We illustrate the approach by applying it to the analytical continuation problem of quantum many-body physics, which involves reconstructing the frequency dependence of physical excitation spectra from data obtained at specific points in the complex frequency plane. Under standard error metrics the method performs as well or better than the Maximum Entropy method for low input noise and is substantially more robust to increased input noise. We expect the methodology to be similarly effective for any problem involving a formally ill-conditioned inversion, provided that the forward problem can be efficiently solved.

Lauren Hannah, David Dunson

Convex regression is a promising area for bridging statistical estimation and deterministic convex optimization. New piecewise linear convex regression methods are fast and scalable, but can have instability when used to approximate constraints or objective functions for optimization. Ensemble methods, like bagging, smearing and random partitioning, can alleviate this problem and maintain the theoretical properties of the underlying estimator. We empirically examine the performance of ensemble methods for prediction and optimization, and then apply them to device modeling and constraint approximation for geometric programming based circuit design.