Fernando submitted his thesis titled 'Embodied Artificial Intelligence: Advanced Deep Reinforcement Learning for Robot Sensorimotor Control' in July 2024 then started a research role in industry as part of the AI Research team at J.P. Morgan, led by Professor Manuela Veloso. The team works on exploring and advancing research in various fields of AI that are impactful to the firm’s clients and its business. Specifically, Fernando works on Optimization, Reinforcement Learning, Decision Making, and Control problems, leveraging the skills he developed during his PhD.
Fernando reports that having undertaken a research internship with the same team as part of his PhD was a positive contribution to being able to take an informed decision on the direction he wanted to follow upon completing. He adds that he is thankful to the CDT for supporting students to undertake internships in industry and remains thankful for all the resources available to CDT students, the community, and the exposure to industrial partners. Finally Fernando adds that he is extremely thankful to his supervisor for having provided a fruitful and supportive environment during his PhD research, and to all the CDT members and staff for their support throughout his journey.
Felix Biggs (Cohort 1)
Felix was awarded a PhD in June 2024 with his thesis 'Exploring Generalisation Performance through PAC-Bayes' and is now a senior machine learning researcher at SecondMind (www.secondmind.ai), working on active learning.
Thesis Abstract: Generalisation in machine learning refers to the ability of a predictor learned on some dataset to perform accurately on new, unseen data. Without generalisation, we might be able to memorise the training data perfectly while predicting poorly on new data, a pathology known as over-fitting. Despite its centrality, the generalisation behaviour of many methods remains poorly understood, particularly in complex domains such as deep learning. Indeed, some models that should over-fit according to traditional theoretical bounds do not. This thesis addresses these issues, particularly in the context of classification, and introduces innovative methods for producing non-vacuous generalisation bounds. The primary thrust is in the development and application of PAC-Bayesian bounds, which are usually used as a method for studying the generalisation of randomised predictors. We begin with an introduction to previous work on the problem of generalisation and to PAC-Bayesian ideas, before applying these in work based on a series of five papers (referenced a-e below). Firstly in (a), we provide lower-variance methods for training stochastic neural networks methods, improving the use of these PAC-Bayes bounds as training objectives. Then, we use PAC-Bayes as a stepping stone to provide non-randomised bounds: (b) using margins, both in general and for several different classifiers; (c) for a specific class of deterministic shallow neural networks (where our bounds are the first to be non-vacuous on real-world data using standard training methods); (d) for majority voting on finite ensembles of classifiers, providing state-of-the-art (and sometimes sharp) guarantees. Lastly in (e), we introduce a PAC-Bayes bound for a modified excess risk, using information about the relative hardness of data examples to reduce variance and tighten a general bound.
 Samuel Cohen (Cohort 1)
Samuel submitted his thesis titled 'Optimal transport on structured spaces: theory and applications' in January 2024 and is preparing for his viva.
 Yue Feng (Cohort 2)
Yue submitted her thesis titled 'Knowledge Enhanced Task-oriented Dialogue Systems' in May 2024 then went on to join the University of Birmingham as an assistant professor.
Dennis submitted a thesis titled 'Semantic Correspondence in Robot Perception' which looks at how establishing visual semantic correspondences, such as similar parts and points of different objects, can be used to transfer robot manipulation skills across different object, and is currently preparing for his viva.
Akbir submitted his thesis titled 'Aligning Super Intelligent Systems via Multi-Agent Curricula' in April 2025 and is currently preparing for his viva
Robert Kirk (Cohort 2) Robert was awarded his PhD in March 2025 with his thesis titled 'Understanding and Evaluating Generalisation for Superhuman AI Systems'
 Linqing Liu (Cohort 2)
Linqing was awarded her PhD in May 2024 for her thesis titled 'Towards Generalized Open Domain Question Answering Systems'.
Thesis Abstract: Generalization remains a paramount yet unresolved challenge for open-domain question answering (ODQA) systems, impeding their capacity to adeptly handle novel queries and responses beyond the confines of their training data. This thesis conducts a comprehensive exploration of ODQA generalization. Wecommence with a meticulous investigation into the underlying challenges. Drawing upon studies on systematic generalization, we introduce and annotate questions according to three categories that measure different levels and kinds of generalization: training set overlap, compositional generalization and novel-entity generalization. When evaluating six popular parametric and non-parametric models, we find non-parametric models demonstrate proficiency with novel entities but encounter difficulties with compositional generalization. Noteworthy correlations emerge, such as a positive association between question pattern frequency and test accuracy, juxtaposed with a strong negative correlation between entity frequency and test accuracy, attributable to closely related distractors. Factors influencing generalization include cascading errors originating from the retrieval component, question pattern frequency, and entity prevalence. Building on these insights, the focus pivots towards the enhancement of passage retrieval. We propose a novel contextual clue sampling strategy using language models to address the vocabulary mismatch challenge in lexical retrieval for ODQA. This two-step method, comprising filtering and fusion, generates a diverse set of query expansion terms, yielding retrieval accuracy similar to dense methods while notably reducing the index size. The subsequent phase concentrates on refining reader models in ODQA throughAbstract 4 f lat minima optimization techniques, incorporating Stochastic Weight Averaging (SWA) and Sharpness Aware Minimization (SAM). Rigorous benchmarking underscores the impact of dataset characteristics and model architecture on optimizer effectiveness, with SAM particularly excelling in Natural Language Processing tasks. The combination of SWA and SAM yields additional gains, underscoring the pivotal role of flatter minimizers in fostering enhanced generalization for reader models in ODQA.
 Augustine Mavor-Parker (Cohort 1)
Augustine submitted his thesis titled 'Towards Robust and Sample Efficient Deep Reinforcement Learning' in September and is preparing for his viva
Luca submitted his thesis 'Neural Surface Representations' in May 2024 and was awarded a PhD in October 2024. More information on my website.
 Antonin Schrab (Cohort 2)
Antonin submitted his thesis titled 'Optimal Kernel Hypothesis Testing' in September 2024 and is preparing for his viva.
 Jas Semrl (Cohort 1)
Jas was awarded a PhD with his thesis titled 'Finite Representations in Relation Algebra' in December 2023.
Thesis Abstract: Binary relations provide a great abstraction for a number of concepts, both in theoretical and applied topics. This is why structures of binary relations have found applications in formal verification, temporal and spatial reasoning in AI, regular language equivalence, sequent calculi, and more. In general, a finite relation algebra cannot be finitely represented. This negatively impacts the feasibility of implementing any of the aforementioned applications based on these structures. Our work focuses on finding large classes of relational structures for which the finite representation property either holds or fails. Furthermore, we examine related topics such as the decidability of the [finite] representation problem and finite axiomatisability. Moreover, we examine the relationship between these properties. We refine Hirsch’s conjecture on which relation algebra reduct signatures have the finite representation property and prove the negative implication of it. Furthermore, we provide a number of results that reveal a possible direction for proving the positive side. We present the first known signature that fails to have a finitely axiomatisable representation class but has the finite representation property. We generalise the results for the undecidability of the representation decision problem and show that semigroups with the Heyting implication fail to have the said problem decidable. We prove and disprove a number of properties for the structures of binary relations with combined operators, motivated by various topics in computer science. Finally, we show a number of results in the area of weakening relation algebras and show the finite weakening representation property for some signatures with their finite representation property open
 Jingwen Wang (Cohort 1)
Jingwen was awarded his PhD in June 2024 with his thesis titled Exploiting Neural Priors in Visual SLAM.
Thesis Abstract: Traditional simultaneous localisation and mapping (SLAM) has shown great performance in camera tracking and geometry reconstruction in various types of environments. However, to enable more advanced applications such as achieving semantic level understanding of the scene, hole-filling, and scene completion in unobserved regions, some (learnt or general) priors need to be applied. In this thesis we aim to explore the incorporation of specific types of prior information in the 3D reconstruction process, such as pre-learnt shape or semantic priors as well as analytical geometric priors. First, in GO-Surf we leverage neural implicit representations with general geometric priors for accurate and fast surface reconstruction from RGB-D sequences. We represent the scene as a multi-level feature grid plus two tiny MLPs decoding the feature into SDF and colour. The training is supervised with rendering and pseudoSDF losses, plus Eikonal and SDF gradient regularization that encourages surface smoothness and hole-filling. GO-Surf can optimize sequences of 1-2K frames in 15-45 minutes, more than 60 times faster than previous MLP-based method, while maintaining on par performance on standard benchmarks. This work is further extended to a full real-time SLAM system named Co-SLAM. Then, in DSP-SLAM we apply pre-learnt shape priors for complete object shape reconstruction. DSP-SLAM builds a rich and accurate joint map of dense 3D objects and sparse landmark points as background. Objects are detected via instance segmentation, and their shape and pose are optimised using category-specific deep shape embeddings as priors, via a novel second order optimization. Our objectaware bundle adjustment builds a pose-graph to jointly optimize camera poses, obAbstract 4 ject locations and feature points. DSP-SLAM can operate at 10 Hz on 3 different input modalities: monocular, stereo, or stereo+LiDAR. Finally, in SeMLaPS we leverage temporal consistency and geometric priors for real-time online semantic mapping. When segmenting a new RGB-D frame, latent feature maps are re-projected from previous frames, which greatly improves 2D segmentation accuracy and temporal consistency. Next, we propose a quasi-planar over-segmentation method that groups raw 3D map elements into segments based on surface normal. A novel 3D CNN then applies post-processing to the labelled mesh at segment level. SeMLaPS achieves state-of-the-art semantic mapping quality and shows better cross-sensor generalization abilities compared to 3D CNNs.
 Changmin Yu (Cohort 1)
Changmin was awarded a PhD in June 2024 with his thesis titled 'Hippocampus-Inspired Representation Learning for Artificial Agents'.
Thesis Abstract: Spatial representations found in the hippocampal formation of freely moving mammals, such as those of grid cells, appear optimal for spatial navigation, and also afford flexible and generalisable non-spatial behaviours. In this thesis, I propose models for learning and representing the structure underlying high-dimensional observation space in artificial agents, drawing inspiration from hippocampal neuroscience. In the first part of the thesis, I study the construction and identification of latent representations. I propose a novel model for grid cell firing based on Fourier analysis of translation-invariant transition dynamics. I show that effects of arbitrary actions can be predicted using a single neural representation and action-dependent weight modulation, and how this model unifies existing models of grid cells based on predictive planning, continuous attractors, and oscillatory interference. Next, I consider the problem of unsupervised learning of the structured latent manifold underlying population neuronal spiking, such that interdependent behavioural variables can be accurately decoded. I propose a novel amortised inference framework such that the recognition networks explicitly parametrise the posterior latent dependency structure, relaxing the full-factorisation assumption. In the second part, I propose representation learning methods inspired by neuroscience and study their application in reinforcement learning. Inspired by the observation of hippocampal “replay” in both temporally forward and backward directions, I show that incorporating temporally backward predictive reconstruction self-supervision into training world models leads to better sample efficiency and stronger generalisability on continuous control tasks. I then propose a novel intrinsic exploration framework under a similar premise, where the intrinsic novelty bonus is constructed based on both prospective and retrospective information. The resulting agents exhibit higher exploration efficiency and ethologically plausible exploration strategies. I conclude by discussing the general implications of learning and utilisation of latent structures in both artificial and biological intelligence, and potential applications of neuralinspired representation learning beyond reinforcement learning.
Jakob Zeitler Jakob submitted his thesis titled 'Elements of Partial Identification:Theory and Applications in December 2024 and is preparing for his viva | |
Close
