Background and Motivation
Modern deep learning models have achieved remarkable success in computer vision and natural language processing. However, they typically produce overconfident predictions and lack reliable mechanisms to quantify uncertainty. This limitation becomes particularly problematic in high-stakes applications, such as healthcare diagnosis, autonomous systems, and scientific discovery.
Bayesian approaches provide a principled framework for modeling uncertainty by capturing posterior distributions over model parameters or predictions. Despite recent progress in approximate Bayesian deep learning (e.g., Monte Carlo dropout, deep ensembles, Laplace approximations, and variational inference), several challenges remain:
-
Scalability: Many Bayesian inference methods are computationally expensive for modern large models.
-
Incomplete Uncertainty Modeling: Most methods focus on single-modal data and fail to account for uncertainty arising from multi-view or multimodal interactions.
-
Distribution Shifts and Missing Modalities: In real-world settings, modalities may be missing or corrupted, making uncertainty estimation unreliable.
-
Calibration Across Modalities: Existing models often produce poorly calibrated uncertainty when integrating multiple modalities.
-
Decision-making under uncertainty: Current frameworks rarely translate uncertainty estimates into robust downstream decisions.
Addressing these issues is critical for building trustworthy multimodal AI systems.
Research Objectives
The goal of this PhD project is to develop scalable Bayesian uncertainty estimation frameworks for single- and multi-view learning that are robust under distribution shift and missing modalities.
The key objectives include:
-
Develop scalable Bayesian deep learning methods for uncertainty estimation in modern neural architectures.
-
Design principled uncertainty modelling frameworks for multi-view/multimodal learning.
-
Model uncertainty propagation across modalities in fusion architectures.
-
Develop robust learning methods under missing modalities and distribution shift.
-
Design uncertainty-aware decision frameworks for downstream tasks.
Expected Contributions
This PhD project is expected to contribute:
-
Scalable Bayesian uncertainty estimation methods for deep neural networks.
-
A principled Bayesian framework for multimodal uncertainty modeling.
-
Robust learning algorithms under missing modalities and distribution shifts.
-
New uncertainty-aware decision frameworks.
-
Open-source toolkits for multimodal uncertainty estimation.
Expected Outcomes
Academic outputs may include publications in:
-
NeurIPS
-
ICLR
-
ICML
-
CVPR / ICCV
-
ACL / EMNLP
-
IEEE TPAMI / JMLR
The project will also produce:
-
open-source implementations
-
benchmark datasets for multimodal uncertainty.
Required knowledge
-
Strong Mathematical Foundation: Demonstrated expertise in Probability and Statistics, Linear Algebra, and Optimisation. A deep understanding of Bayesian Inference (e.g., Variational Inference, Monte Carlo methods) is highly preferred.
-
Machine Learning Expertise: Mastery of Deep Learning fundamentals, including experience with Vision Transformers (ViTs), Generative Models (e.g., Diffusion or VAEs), and Self-Supervised Learning.
-
Information Theory: Familiarity with concepts like Entropy, Mutual Information, and Kullback-Leibler (KL) Divergence, which are critical for designing acquisition functions.
-
Deep Learning Frameworks: Advanced proficiency in PyTorch. The candidate must be able to implement custom training loops, handle complex data pipelines, and modify model architectures.
-
Software Engineering: Strong Python programming skills, including experience with version control (Git) and writing clean, reproducible, and modular code.
