Semantic segmentation of satellite imagery is crucial for Earth observation applications, but remains constrained by limited labelled training data. While self-supervised pretraining methods like Masked Autoencoders (MAE) have shown promise, they focus on reconstruction rather than localisation-a fundamental aspect of segmentation tasks. We propose adapting LOCA (Location-aware), a position prediction self-supervised learning method, for multimodal satellite imagery semantic segmentation. Our approach addresses the unique challenges of satellite data by extending SatMAE's channel grouping from multispectral to multimodal data, enabling effective handling of multiple modalities, and introducing same-group attention masking to encourage cross-modal interaction during pretraining. The method uses relative patch position prediction, encouraging spatial reasoning for localisation rather than reconstruction. We evaluate our approach on the Sen1Floods11 flood mapping dataset, where it significantly outperforms existing reconstruction-based self-supervised learning methods for satellite imagery. Our results demonstrate that position prediction tasks, when properly adapted for multimodal satellite imagery, learn representations more effective for satellite image semantic segmentation than reconstruction-based approaches.

This thesis addresses challenges related to data and parameter efficiency in neural language models, with a focus on representation analysis and the introduction of new optimization techniques. The first part examines the properties and dynamics of language representations within neural models, emphasizing their significance in enhancing robustness and generalization. It proposes innovative approaches based on representation smoothness, including regularization strategies that utilize Jacobian and Hessian matrices to stabilize training and mitigate sensitivity to input perturbations. The second part focuses on methods to significantly enhance data and parameter efficiency by integrating active learning strategies with parameter-efficient fine-tuning, guided by insights from representation smoothness analysis. It presents smoothness-informed early-stopping techniques designed to eliminate the need for labeled validation sets and proposes innovative combinations of active learning and parameter-efficient fine-tuning to reduce labeling efforts and computational resources. Extensive experimental evaluations across various NLP tasks demonstrate that these combined approaches substantially outperform traditional methods in terms of performance, stability, and efficiency. The third part explores weak supervision techniques enhanced by in-context learning to effectively utilize unlabeled data, further reducing dependence on extensive labeling. It shows that using in-context learning as a mechanism for weak supervision enables models to better generalize from limited labeled data by leveraging unlabeled examples more effectively during training. Comprehensive empirical evaluations confirm significant gains in model accuracy, adaptability, and robustness, especially in low-resource settings and dynamic data environments.

Self-supervised learning (SSL) is able to build latent representations that generalize well to unseen data. However, only a few SSL techniques exist for the online CL setting, where data arrives in small minibatches, the model must comply with a fixed computational budget, and task boundaries are absent. We introduce Continual Latent Alignment (CLA), a novel SSL strategy for Online CL that aligns the representations learned by the current model with past representations to mitigate forgetting. We found that our CLA is able to speed up the convergence of the training process in the online scenario, outperforming state-of-the-art approaches under the same computational budget. Surprisingly, we also discovered that using CLA as a pretraining protocol in the early stages of pretraining leads to a better final performance when compared to a full i.i.d. pretraining.

Tabular data synthesis for supervised learning ('SL') model training is gaining popularity in industries such as healthcare, finance, and retail. Despite the progress made in tabular data generators, models trained with synthetic data often underperform compared to those trained with original data. This low SL utility of synthetic data stems from class imbalance exaggeration and SL data relationship overlooked by tabular generator. To address these challenges, we draw inspirations from techniques in emerging data-centric artificial intelligence and elucidate Pruning and ReOrdering ('PRRO'), a novel pipeline that integrates data-centric techniques into tabular data synthesis. PRRO incorporates data pruning to guide the table generator towards observations with high signal-to-noise ratio, ensuring that the class distribution of synthetic data closely matches that of the original data. Besides, PRRO employs a column reordering algorithm to align the data modeling structure of generators with that of SL models. These two modules enable PRRO to optimize SL utility of synthetic data. Empirical experiments on 22 public datasets show that synthetic data generated using PRRO enhances predictive performance compared to data generated without PRRO. Specifically, synthetic replacement of original data yields an average improvement of 26.74% and up to 871.46% improvement using PRRO, while synthetic appendant to original data results with PRRO-generated data results in an average improvement of 6.13% and up to 200.32%. Furthermore, experiments on six highly imbalanced datasets show that PRRO enables the generator to produce synthetic data with a class distribution that resembles the original data more closely, achieving a similarity improvement of 43%. Through PRRO, we foster a seamless integration of data synthesis to subsequent SL prediction, promoting quality and accessible data analysis.

Label Proportion Learning (LLP) addresses the classification problem where multiple instances are grouped into bags and each bag contains information about the proportion of each class. However, in practical applications, obtaining precise supervisory information regarding the proportion of instances in a specific class is challenging. To better align with real-world application scenarios and effectively leverage the proportional constraints of instances within tuples, this paper proposes a generalized learning framework \emph{MDPU}. Specifically, we first mathematically model the distribution of instances within tuples of arbitrary size, under the constraint that the number of positive instances is no less than that of negative instances. Then we derive an unbiased risk estimator that satisfies risk consistency based on the empirical risk minimization (ERM) method. To mitigate the inevitable overfitting issue during training, a risk correction method is introduced, leading to the development of a corrected risk estimator. The generalization error bounds of the unbiased risk estimator theoretically demonstrate the consistency of the proposed method. Extensive experiments on multiple datasets and comparisons with other relevant baseline methods comprehensively validate the effectiveness of the proposed learning framework.

Camera traps are revolutionising wildlife monitoring by capturing vast amounts of visual data; however, the manual identification of individual animals remains a significant bottleneck. This study introduces a fully self-supervised approach to learning robust chimpanzee face embeddings from unlabeled camera-trap footage. Leveraging the DINOv2 framework, we train Vision Transformers on automatically mined face crops, eliminating the need for identity labels. Our method demonstrates strong open-set re-identification performance, surpassing supervised baselines on challenging benchmarks such as Bossou, despite utilising no labelled data during training. This work underscores the potential of self-supervised learning in biodiversity monitoring and paves the way for scalable, non-invasive population studies.

Curriculum learning is a widely adopted training strategy in natural language processing (NLP), where models are exposed to examples organized by increasing difficulty to enhance learning efficiency and performance. However, most existing approaches rely on manually defined difficulty metrics -- such as text length -- which may not accurately reflect the model's own perspective. To overcome this limitation, we present a self-adaptive curriculum learning paradigm that prioritizes fine-tuning examples based on difficulty scores predicted by pre-trained language models (PLMs) themselves. Building on these scores, we explore various training strategies that differ in the ordering of examples for the fine-tuning: from easy-to-hard, hard-to-easy, to mixed sampling. We evaluate our method on four natural language understanding (NLU) datasets covering both binary and multi-class classification tasks. Experimental results show that our approach leads to faster convergence and improved performance compared to standard random sampling.

To alleviate the annotation burden in supervised learning, N-tuples learning has recently emerged as a powerful weakly-supervised method. While existing N-tuples learning approaches extend pairwise learning to higher-order comparisons and accommodate various real-world scenarios, they often rely on task-specific designs and lack a unified theoretical foundation. In this paper, we propose a general N-tuples learning framework based on empirical risk minimization, which systematically integrates pointwise unlabeled data to enhance learning performance. This paper first unifies the data generation processes of N-tuples and pointwise unlabeled data under a shared probabilistic formulation. Based on this unified view, we derive an unbiased empirical risk estimator that generalizes a broad class of existing N-tuples models. We further establish a generalization error bound for theoretical support. To demonstrate the flexibility of the framework, we instantiate it in four representative weakly supervised scenarios, each recoverable as a special case of our general model. Additionally, to address overfitting issues arising from negative risk terms, we adopt correction functions to adjust the empirical risk. Extensive experiments on benchmark datasets validate the effectiveness of the proposed framework and demonstrate that leveraging pointwise unlabeled data consistently improves generalization across various N-tuples learning tasks.

Contrastive learning (CL) has recently emerged as an alternative to traditional supervised machine learning solutions by enabling rich representations from unstructured and unlabeled data. However, CL and, more broadly, self-supervised learning (SSL) methods often demand a large amount of data and computational resources, posing challenges for deployment on resource-constrained edge devices. In this work, we explore the feasibility and efficiency of SSL techniques for edge-based learning, focusing on trade-offs between model performance and energy efficiency. In particular, we analyze how different SSL techniques adapt to limited computational, data, and energy budgets, evaluating their effectiveness in learning robust representations under resource-constrained settings. Moreover, we also consider the energy costs involved in labeling data and assess how semi-supervised learning may assist in reducing the overall energy consumed to train CL models. Through extensive experiments, we demonstrate that tailored SSL strategies can achieve competitive performance while reducing resource consumption by up to 4X, underscoring their potential for energy-efficient learning at the edge.

Interpretability research often aims to predict how a model will respond to targeted interventions on specific mechanisms. However, it rarely predicts how a model will respond to unseen input data. This paper explores the promises and challenges of interpretability as a tool for predicting out-of-distribution (OOD) model behavior. Specifically, we investigate the correspondence between attention patterns and OOD generalization in hundreds of Transformer models independently trained on a synthetic classification task. These models exhibit several distinct systematic generalization rules OOD, forming a diverse population for correlational analysis. In this setting, we find that simple observational tools from interpretability can predict OOD performance. In particular, when in-distribution attention exhibits hierarchical patterns, the model is likely to generalize hierarchically on OOD data -- even when the rule's implementation does not rely on these hierarchical patterns, according to ablation tests. Our findings offer a proof-of-concept to motivate further interpretability work on predicting unseen model behavior.