Publications

Knowledge distillation is considered a compression mechanism when judged on the resulting student's accuracy and loss, yet its functional impact is poorly understood. We quantify the compression capacity of knowledge distillation and the resulting knowledge transfer from a functional perspective, decoupling compression from architectural reduction to provide an improved understanding of knowledge distillation. We employ a control-driven experimental protocol with hypothesis testing and random control distillation to isolate and understand knowledge transfer mechanisms across data modalities. To test the breadth and limits of our analyses, we study self-distillation, standard distillation, feature-map matching variants, distillation capacity scaling across model sizes, and the impact of temperature on knowledge transfer. We find statistically supported knowledge transfer in some modalities and architectures; however, the extent of this transfer is less pronounced than anticipated, even under conditions that maximise knowledge sharing. Notably, in cases of significant functional transfer, we identify a consistent and severe asymmetric transfer of negative knowledge to the student, raising safety concerns for knowledge distillation. Across 22 experimental setups, 9 architectures, and 7 datasets, our results suggest that knowledge distillation functions less as a robust compression-by-transfer mechanism and more as a data-dependent regulariser whose transfer component is biased towards negative asymmetric transfer.

Flat minima are strongly associated with improved generalisation in deep neural networks. However, this connection has proven nuanced in recent studies, with both theoretical counterexamples and empirical exceptions emerging in the literature. In this paper, we revisit the role of sharpness in model performance and examine sharpness as a function-dependent property rather than as a direct indicator of poor generalisation. We conduct extensive empirical studies ranging from single-objective optimisation, synthetic non-linear binary classification tasks, to modern image classification tasks. In single-objective optimisation, we show that flatness and sharpness are relative to the function being learned: equally optimal solutions can exhibit markedly different local geometry. In synthetic non-linear binary classification tasks, we show that increasing decision-boundary tightness can increase sharpness even when models generalise perfectly, suggesting that sharpness is not reducible to memorisation alone. Finally, in large-scale experiments, we find that sharper minima often emerge when models are regularised (e.g., via weight decay, data augmentation, or SAM), and can coincide with better generalisation, calibration, robustness, and functional consistency. Our findings are consistent with a function-centric interpretation of minima geometry: flatness should not be treated as a universal proxy for generalisation, and sharper minima can reflect desirable inductive biases associated with improved predictive behaviour, calling for a function-centric reappraisal of minima geometry.

Data-free methods for analysing and understanding the layers of neural networks offer many metrics for quantifying notions of strong' versus weak' layers, with the promise of increased interpretability. In particular, random matrix theory (RMT) offers data-free metrics that claim predictive power over the quality of pre-trained models at a layerwise precision and the ability to identify model pathologies, indicating a unique relationship with generalisation. As a result, metrics from RMT are championed for pre-training optimisation strategies and post-training compression. We establish that RMT-based metrics are unrelated to performance or training by showing that functionally indistinguishable reparametrisations of a pre-trained model can have arbitrary metrics. We show this across a range of architectures and scales. To create functionally indistinguishable reparametrisations, we exploit the well-established phenomenon of criticality: some layers can be re-initialised or re-randomised without affecting the functional behaviour of the model -- they are called robust -- while others cannot -- they are called critical. Re-initialising or re-randomising robust layers provides functionally indistinguishable reparametrisations for which RMT-based metrics are arbitrary. Moreover, we show that relationships between metrics offered by RMT are spuriously related to generalisation. We conclude by showing that if many pre-trained models have data-free metrics in a `good' range, it is, in part, dependent on model initialisation.

Sound design workflows frequently oscillate between time-consuming library searches and the complexity of procedural synthesis, with practitioners typically relying on disconnected tools to address each challenge separately. This paper introduces Quality Audio Pro- totyping (QuAP), a working prototype that unifies content-based audio retrieval and procedural sound generation within a single in- terface, reducing the procedural distance between a narrative con- cept and its sonic realisation. QuAP integrates a similarity-based retrieval engine with real-time procedural audio models, comple- mented by a rule-based assistant that provides perceptually in- formed parameter guidance, offering definitions and recommen- dations derived from empirical optimisation rather than requiring prior synthesis knowledge. Preliminary evaluation confirms the vi- ability of this approach: subjective assessment demonstrated sta- tistically significant quality improvements in five of six embedded synthesis models, and an encoder ablation study established the preferred retrieval architecture on a sound effect dataset. A user evaluation with 16 practitioners confirmed the tool’s workflow util- ity, with all participants agreeing that the parameter assistant pre- served creative agency while lowering the barrier to procedural interaction.

The geometric flatness of neural network minima has long been associated with desirable generalisation properties. In this paper, we extensively explore the hypothesis that robust, calibrated and functionally similar models sit at flatter minima, inline with prevailing understandings of the relationship between flatness and generalisation. Contrary to common assertions in the literature, we find a relationship between increased sharpness, generalisaton, calibration and robustness in neural networks across architectures when using Sharpness Aware Minimisation, augmentation and weight decay as regulariser controls. Our findings suggest that the role of increased sharpness should be considered independently for individual models when reasoning about the geometric properties of neural networks. We show that sharpness can be related to generalisation and safety-relevant properties against the flatter minima found without the use of our regularisation controls. Understanding these properties calls for a re-thinking of the role of sharpness in geometric landscapes.

Data-free methods for analysing and understanding the layers of neural networks have offered many metrics for quantifying notions of strong" versus weak" layers, with the promise of increased interpretability. We examine how robust data-free metrics are under random control conditions of critical and robust layers. Contrary to the literature, we find counter-examples that provide counter-evidence to the efficacy of data-free methods. We show that data-free metrics are not reparameterisation invariant in these conditions and lose predictive capacity across correlation measures, RMSE, Person Coefficient and Kendall's Tau measure. Thus, we argue that to understand neural networks fundamentally, we must rigorously analyse the interactions between data, weights, and resulting functions that contribute to their outputs -- contrary to traditional Random Matrix Theory perspectives.

Grokking is a delayed transition from memorisation to generalisation in neural networks. It challenges efficient learning, particularly in structured tasks and small-data regimes. We explore grokking in modular arithmetic from the perspective of a training pathology. We use Singular Value Decomposition (SVD) to modify the weight matrices of neural networks by changing the representation of the weight matrix, $W$, into the product of three matrices, $U$, $\Sigma$ and $V^T$ . Through empirical evaluations on the modular addition task, we show that this representation significantly reduces the effect of grokking and, in some cases, eliminates it.

AI Policymakers are responsible for delivering effective governance mechanisms that can provide safe, aligned and trustworthy AI development. However, the information environment offered to policymakers is characterized by an unnecessarily low signal-to-noise ratio, favouring regulatory capture and creating deep uncertainty and divides on which risks should be prioritized from a governance perspective. We posit that the current speed of publication in AI combined with the lack of strong scientific standards, via weak reproducibility protocols, effectively erodes the power of policymakers to enact meaningful policy and governance protocols. Our paper outlines how AI research could adopt stricter reproducibility guidelines to assist governance endeavours and improve consensus on the risk landscapes posed by AI. We evaluate the forthcoming reproducibility crisis within AI research through the lens of reproducibility crises in other scientific domains and provide a commentary on how adopting preregistration, increased statistical power and negative result publication reproducibility protocols can enable effective AI governance. While we maintain that AI governance must be reactive due to AI's significant societal implications we argue that policymakers and governments must consider reproducibility protocols as a core tool in the governance arsenal and demand higher standards for AI research.

Empirical findings of accuracy correlations between students and teachers in the knowledge distillation framework have served as supporting evidence for knowledge transfer. In this paper, we sought to explain and understand the knowledge transfer derived from knowledge distillation via functional similarity, hypothesising that knowledge distillation provides a functionally similar student to its teacher model. While we accept this hypothesis for two out of three architectures across a range of metrics for functional analysis against four controls, the results show that knowledge transfer is significant but it is less pronounced than expected for conditions that maximise opportunities for functional similarity. Furthermore, results from the use of Uniform and Gaussian Noise as teachers suggest that the knowledge-sharing aspects of knowledge distillation inadequately describe the accuracy benefits witnessed when using the knowledge distillation training setup itself. Moreover, in the first instance, we show that knowledge distillation is not a compression mechanism but primarily a data-dependent training regulariser with a small capacity to transfer knowledge in the best case.

We probe the relation between flatness, generalisation and calibration in neural networks, using explicit regularisation as a control variable. Our findings indicate that the range of flatness metrics surveyed fail to positively correlate with variation in generalisation or calibration. In fact, the correlation is often opposite to what has been hypothesized or claimed in prior work, with calibrated models typically existing at sharper minima compared to relative baselines, this relation exists across model classes and dataset complexities.

Compression techniques, such as Knowledge distillation, Pruning, and Quantiza- tion reduce the computational costs of model inference and enable on-edge ma- chine learning. The efficacy of compression methods is often evaluated through the proxy of accuracy and loss to understand similarity of the compressed model. This study aims to explore the functional divergence between compressed and uncompressed models. The results indicate that Quantization and Pruning create models that are functionally similar to the original model. In contrast, Knowl- edge distillation creates models that do not functionally approximate their teacher models. The compressed model resembles the dissimilarity of function observed in independently trained models. Therefore, it is verified, via a functional under- standing, that Knowledge distillation is not a compression method. Thus, leading to the definition of Knowledge distillation as a training regulariser given that no knowledge is distilled from a teacher to a student.