Institution profile

Chalmers University of Technology

Academic institutioneurope · se
Official website
Research library532linked papers
Opportunities0open roles
Selected work

Representative Papers

MCMC-Correction of Score-Based Diffusion Models for Model Composition

Jul 26, 2023arXiv.org

While energy-parameterized diffusion models support Metropolis-Hastings (MH)-based MCMC sampling—yielding substantially improved sample quality under model composition—standard score-parameterized models lack an explicit energy function, preventing direct MH correction. Method: We propose Score-MCMC, a framework that reconstructs energy differences via line-integral approximations of the pre-trained score function, enabling derivation of a computable MH acceptance probability without modifying the underlying score model. Contribution/Results: Score-MCMC theoretically and practically bridges score-based and energy-based MCMC sampling. By coupling the diffusion reverse process with MH correction, it achieves sampling fidelity and diversity on par with energy-parameterized models across multiple benchmarks—particularly enhancing compositional distribution modeling in terms of both accuracy and sample diversity.

3 citationsRead paper

A Theory of Initialisation's Impact on Specialisation

Mar 04, 2025

This work challenges the necessity of neuron specialization for mitigating catastrophic forgetting in continual learning, revealing that specialization is primarily governed by network initialization rather than intrinsic task properties. Method: Through theoretical analysis and empirical validation, the authors demonstrate that weight imbalance and high weight entropy actively induce localized representations; they provide the first theoretical proof that specialization is not inherent but contingent. They further derive a quantitative relationship between specialization degree and initialization parameters, and reproduce the monotonic relationship between task similarity and forgetting rate even in non-specialized networks. Contribution/Results: Specialized initialization significantly enhances Elastic Weight Consolidation (EWC) performance, an effect attributable to initialization-induced prior shaping of representation structure. These findings establish a novel theoretical foundation for regularization design in continual learning and yield principled guidelines for initialization strategy selection.

1 citationsRead paper

Automating Boundary Filling in Cubical Agda

Feb 19, 2024International Conference on Formal Structures for Computation and Deduction

This work addresses the challenge of automating proofs involving higher-dimensional homotopical structures in cubical type theory, focusing on boundary filling—the core task underlying both contortion solving (single-cube filling) and Kan solving (multi-cube gluing). Methodologically, it models contortion as a poset-mapping problem and introduces, for the first time, a constraint satisfaction programming (CSP) framework to solve Kan filling. To mitigate combinatorial explosion in high dimensions and undecidability inherent in the theory, it designs lightweight heuristic algorithms. The prototype solver, implemented in Haskell, integrates poset-based modeling, CSP solving, and an interface to Cubical Agda. It successfully automates the verification of the Eckmann–Hilton theorem and demonstrates efficiency across multiple benchmark suites. This work establishes a novel paradigm and delivers key technical foundations for practical proof automation in cubical type theory.

1 citationsRead paper

Compositionality in algorithms for smoothing

Mar 24, 2023

This paper addresses the lack of compositional semantics in probabilistic algorithms—particularly the bilateral filter-based Gaussian (BFFG) algorithm. To resolve this, the authors establish, for the first time, a rigorous connection between BFFG and optics in category theory: they model BFFG as a functor from the category of Markov kernels to the category of optics and prove that this functor carries a lax monad structure. This categorical construction exposes BFFG’s intrinsic compositional mechanism and provides it with a principled, category-theoretic semantic foundation. The key contribution lies in elevating classical stochastic algorithms to higher-order abstractions endowed with algebraic structure—enabling modular design, formal verification, and cross-model reuse. By unifying probabilistic computation with optic-based compositional principles, the work opens a new theoretical pathway for compositional reasoning about probabilistic programs and randomized algorithms.

1 citationsRead paper
Recent publications

Latest Papers