Marcin B. Tomczak

Marcin B. Tomczak

University of Cambridge
1 previous affiliation
3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Marcin B. Tomczak focuses on the development of scalable Bayesian inference techniques for deep learning, with a specific emphasis on improving the robustness, optimization, and predictive performance of Bayesian Neural Networks (BNNs). His work addresses the computational bottlenecks and underfitting issues that have traditionally limited the practical application of BNNs in large-scale machine learning tasks. By refining variational inference (VI) frameworks and exploring novel reparameterization strategies, he seeks to provide more reliable uncertainty quantification and automated hyperparameter tuning in complex neural architectures.

Research Trajectory

The evolution of Tomczak’s research reflects a transition from addressing the fundamental instabilities of reinforcement learning (RL) to developing more rigorous probabilistic frameworks for deep neural networks. In the earlier stages of his observed research activity around 2019, his work was closely aligned with the challenges of Robust Optimization and Stability in Deep Reinforcement Learning. This area of inquiry focused on the inherent fragility of deep RL agents, which are often sensitive to random seeds and hyperparameter configurations. A key component of this research involved investigating Trust Region and Divergence Regularization for Stable Policy Optimization, a methodology designed to prevent the "catastrophic collapse" of learning by constraining policy updates within a safe divergence threshold.

During this period, Tomczak also engaged with the complexities of long-term dependency in sequential decision-making. This included work on Temporal Credit Assignment for Long Horizon Decisions, which addresses the difficulty of attributing rewards to specific actions across extended time scales. This line of research is particularly relevant to Breaking the Curse of Horizon in Off Policy Evaluation, where the goal is to predict the performance of new strategies using historical data—a critical requirement for high-stakes domains such as autonomous driving or healthcare where real-world testing of unverified policies is prohibited.

Beginning in 2020, Tomczak’s focus shifted toward the optimization and scalability of Bayesian Neural Networks. This transition was marked by a move away from the "recency bias" and stability issues of RL toward the structural pathologies of variational inference in BNNs. He identified that standard mean-field variational inference, which assumes a fully factorized distribution, often fails to capture the highly correlated parameters characteristic of deep networks. His subsequent work has focused on deriving more expressive approximating families and tighter evidence lower bounds (ELBOs) to mitigate underfitting and improve the reliability of Bayesian deep learning.

Current Directions

Tomczak’s current research is centered on three primary pillars of Bayesian deep learning: computational efficiency, hyperparameter inference, and prior sensitivity. A major ongoing challenge in this field is the "underfitting" problem, where BNNs often perform worse than their deterministic counterparts. Tomczak addresses this by developing novel variational bounds that treat prior parameters distributionally rather than as fixed values. This approach allows for the simultaneous learning of weights and the hyperparameters that govern their distribution, leading to models that are more resilient to the choice of initial priors.

Furthermore, he is investigating methods to scale Gaussian variational inference to high-dimensional models. While Gaussian approximations can capture the correlations between weights that mean-field methods ignore, they are typically computationally prohibitive for large networks. Tomczak’s work in low-rank approximations and new reparameterization tricks aims to make these sophisticated posterior approximations viable for modern deep learning architectures. By combining local reparameterization with importance sampling, he is also exploring more direct ways to estimate the gradient of the marginal log-likelihood, which serves as a more principled objective for model selection and hyperparameter optimization than standard cross-validation.

Representative Work

In Efficient Low Rank Gaussian Variational Inference for Neural Networks, Tomczak (as first author) addresses the scalability limitations of Gaussian approximating families in variational inference. While mean-field distributions are computationally cheap, they are often insufficient for BNNs due to their inability to model parameter correlations. Tomczak derives a computationally efficient algorithm that utilizes a low-rank plus diagonal covariance structure. By introducing a new form of the reparameterization trick, the method reduces the variance of gradient updates and allows Gaussian VI to scale to larger models, providing a more expressive alternative to standard factorized distributions without the prohibitive cost of full-rank matrices.

Continuing the focus on BNN optimization, Marginal Likelihood Gradient for Bayesian Neural Networks explores the use of the marginal log-likelihood (MLL) as a tool for hyperparameter inference. Standard variational methods often struggle to provide accurate MLL estimates. In this work, Tomczak and his collaborator derive a practical estimator for the gradient of the MLL by combining the local reparameterization of the network with a self-normalized importance sampling estimator. This approach allows for the automatic inference of important hyperparameters, protecting against over-fitting and reducing the reliance on manual tuning or computationally expensive cross-validation.

In Collapsed Variational Bounds for Bayesian Neural Networks, Tomczak (as first author) tackles the sensitivity of BNNs to prior weight distributions, a common cause of underfitting. The paper proposes an extension of the standard BNN model that treats prior parameters as random variables. By analytically collapsing the variational bound with respect to the posteriors of these prior parameters, Tomczak derives novel and tighter ELBOs. The experimental results demonstrate that these collapsed bounds significantly improve predictive performance and reduce the sensitivity of the model to the initial choice of prior, addressing one of the most significant barriers to the adoption of Bayesian methods in deep learning.

While the evidence for Tomczak's contributions to reinforcement learning is primarily situated within broader thematic clusters, his first-author publications in 2020 and 2021 establish a clear leadership role in the field of Bayesian optimization. His work consistently seeks to bridge the gap between the theoretical advantages of Bayesian inference—such as principled uncertainty and automated model selection—and the practical requirements of modern, high-dimensional neural network training.

Profile updated

See something wrong? Report a mistake