Papers

Asks whether a spoken word classifier can sequentially learn to distinguish 1000 classes from only five examples per class. We train a model with the Generative Meta-Continual Learning (GeMCL) algorithm and compare it against strong baselines for this large-scale, few-shot continual setting.

GeMCL delivers exceptionally stable performance and adapts roughly 2000× faster than a frozen HuBERT encoder with repeatedly retrained classifiers, though it does not always surpass fully fine-tuned baselines. The result points toward practical, rapidly-adaptable keyword systems that keep accumulating new classes without retraining from scratch.

Meta-learning outperforms supervised learning for few-shot monolingual spoken word classification, yet remains under-explored in the multilingual setting. We apply the Generative Meta-Continual Learning (GeMCL) algorithm to spoken word classification — its generative nature makes it viable in application, while the meta-learning component promotes the generalisation that multilingual use demands.

We train monolingual models on English, German, French, and Catalan, a bilingual model on English and German, and a multilingual model on all four languages. Although the multilingual model performs best, the differences between models are unexpectedly small. We also find that the hours of unique data seen during training is a stronger performance indicator than the number of languages in the training data.

Investigates how feature resolution — the constraint that each data point exposes only a subset of the available features — affects the embedding dimension needed to identify individual features. Using feature composition and poset theory, we derive upper bounds on the dimensionality required for reliable feature identification.

We show that even a slight uniform dependence between features can be exploited to reduce the required embedding dimension by at least a third while still keeping individual features identifiable. Published as a peer-reviewed post in the ICLR 2026 Blog Track, with an accompanying poster.

Proposes MAMLCon, an extension of model-agnostic meta-learning (MAML) for continual few-shot spoken word classification. Each inner learning loop concludes with a consolidation gradient step that uses stored templates (one per class) drawn from all previously seen classes, letting the model rehearse without a large replay buffer.

MAMLCon consistently outperforms OML (Online-aware Meta-Learning) across varying numbers of shots and final class counts on Google Speech Commands and the FACC dataset. It targets the practical setting of user-defined keyword systems where new words are added incrementally, showing that meta-learning can mitigate catastrophic forgetting in genuinely on-device regimes.

Introduces the Representation Manifold Quality Metric (RMQM), a principled geometric measure of learned representation quality that correlates positively with downstream task performance.

We characterise representation manifolds by tracking how data points move under sequentially larger perturbations — white-noise injection and PGD adversarial attacks — revealing that self-supervised methods learn smoother manifolds with larger but more consistent step sizes. RMQM gives a framework for comparing pretraining methods beyond the usual linear-probe accuracy, enabling more detailed structural analysis of embedding spaces.