The federated learning paradigm has motivated the development of methods for aggregating multiple client updates into a global server model, without sharing client data. Many federated learning algorithms, including the canonical Federated Averaging (FedAvg), take a direct (possibly weighted) average of the client parameter updates, motivated by results in distributed optimization. In this work, we adopt a function space perspective and propose a new algorithm, FedFish, that aggregates local approximations to the functions learned by clients, using an estimate based on their Fisher information. We evaluate FedFish on realistic, large-scale cross-device benchmarks. While the performance of FedAvg can suffer as client models drift further apart, we demonstrate that FedFish is more robust to longer local training. Our evaluation across several settings in image and language benchmarks shows that FedFish outperforms FedAvg as local training epochs increase. Further, Fed-Fish results in global networks that are more amenable to efficient personalization via local fine-tuning on the same or shifted data distributions. For instance, federated pretraining on the C4 dataset, followed by few-shot personalization on Stack Overflow, results in a 7% improvement in next-token prediction by FedFish over FedAvg.

Screening is critical for prevention and early detection of cervical cancer but it is time-consuming and laborious. Supervised deep convolutional neural networks have been developed to automate pap smear screening and the results are promising. However, the interest in using only normal samples to train deep neural networks has increased owing to class imbalance problems and high-labeling costs that are both prevalent in healthcare. In this study, we introduce a method to learn explainable deep cervical cell representations for pap smear cytology images based on one class classification using variational autoencoders. Findings demonstrate that a score can be calculated for cell abnormality without training models with abnormal samples and localize abnormality to interpret our results with a novel metric based on absolute difference in cross entropy in agglomerative clustering. The best model that discriminates squamous cell carcinoma (SCC) from normals gives 0.908 +- 0.003 area under operating characteristic curve (AUC) and one that discriminates high-grade epithelial lesion (HSIL) 0.920 +- 0.002 AUC. Compared to other clustering methods, our method enhances the V-measure and yields higher homogeneity scores, which more effectively isolate different abnormality regions, aiding in the interpretation of our results. Evaluation using in-house and additional open dataset show that our model can discriminate abnormality without the need of additional training of deep models.

With the ever-growing presence of social media platforms comes the increased spread of harmful content and the need for robust hate speech detection systems. Such systems easily overfit to specific targets and keywords, and evaluating them without considering distribution shifts that might occur between train and test data overestimates their benefit. We challenge hate speech models via new train-test splits of existing datasets that rely on the clustering of models' hidden representations. We present two split variants (Subset-Sum-Split and Closest-Split) that, when applied to two datasets using four pretrained models, reveal how models catastrophically fail on blind spots in the latent space. This result generalises when developing a split with one model and evaluating it on another. Our analysis suggests that there is no clear surface-level property of the data split that correlates with the decreased performance, which underscores that task difficulty is not always humanly interpretable. We recommend incorporating latent feature-based splits in model development and release two splits via the GenBench benchmark.

Machine learning assumes a pivotal role in our data-driven world. The increasing scale of models and datasets necessitates quick and reliable algorithms for model training. This dissertation investigates adaptivity in machine learning optimizers. The ensuing chapters are dedicated to various facets of adaptivity, including: 1. personalization and user-specific models via personalized loss, 2. provable post-training model adaptations via meta-learning, 3. learning unknown hyperparameters in real time via hyperparameter variance reduction, 4. fast O(1/k^2) global convergence of second-order methods via stepsized Newton method regardless of the initialization and choice basis, 5. fast and scalable second-order methods via low-dimensional updates. This thesis contributes novel insights, introduces new algorithms with improved convergence guarantees, and improves analyses of popular practical algorithms.

This paper considers how interactions with AI algorithms can boost human creative thought. We employ a psychological task that demonstrates limits on human creativity, namely semantic feature generation: given a concept name, respondents must list as many of its features as possible. Human participants typically produce only a fraction of the features they know before getting "stuck." In experiments with humans and with a language AI (GPT-4) we contrast behavior in the standard task versus a variant in which participants can ask for algorithmically-generated hints. Algorithm choice is administered by a multi-armed bandit whose reward indicates whether the hint helped generating more features. Humans and the AI show similar benefits from hints, and remarkably, bandits learning from AI responses prefer the same prompting strategy as those learning from human behavior. The results suggest that strategies for boosting human creativity via computer interactions can be learned by bandits run on groups of simulated participants.

Despite the prevalence of tabular datasets, few-shot learning remains under-explored within this domain. Existing few-shot methods are not directly applicable to tabular datasets due to varying column relationships, meanings, and permutational invariance. To address these challenges, we propose FLAT-a novel approach to tabular few-shot learning, encompassing knowledge sharing between datasets with heterogeneous feature spaces. Utilizing an encoder inspired by Dataset2Vec, FLAT learns low-dimensional embeddings of datasets and their individual columns, which facilitate knowledge transfer and generalization to previously unseen datasets. A decoder network parametrizes the predictive target network, implemented as a Graph Attention Network, to accommodate the heterogeneous nature of tabular datasets. Experiments on a diverse collection of 118 UCI datasets demonstrate FLAT's successful generalization to new tabular datasets and a considerable improvement over the baselines.

Interpretability in machine learning is critical for the safe deployment of learned policies across legally-regulated and safety-critical domains. While gradient-based approaches in reinforcement learning have achieved tremendous success in learning policies for continuous control problems such as robotics and autonomous driving, the lack of interpretability is a fundamental barrier to adoption. We propose Interpretable Continuous Control Trees (ICCTs), a tree-based model that can be optimized via modern, gradient-based, reinforcement learning approaches to produce high-performing, interpretable policies. The key to our approach is a procedure for allowing direct optimization in a sparse decision-tree-like representation. We validate ICCTs against baselines across six domains, showing that ICCTs are capable of learning policies that parity or outperform baselines by up to 33% in autonomous driving scenarios while achieving a 300x-600x reduction in the number of parameters against deep learning baselines. We prove that ICCTs can serve as universal function approximators and display analytically that ICCTs can be verified in linear time. Furthermore, we deploy ICCTs in two realistic driving domains, based on interstate Highway-94 and 280 in the US. Finally, we verify ICCT's utility with end-users and find that ICCTs are rated easier to simulate, quicker to validate, and more interpretable than neural networks.

A novel approach is presented for discovering PDEs that govern the motion of satellites in space. The method is based on SINDy, a data-driven technique capable of identifying the underlying dynamics of complex physical systems from time series data. SINDy is utilized to uncover PDEs that describe the laws of physics in space, which are non-deterministic and influenced by various factors such as drag or the reference area (related to the attitude of the satellite). In contrast to prior works, the physically interpretable coordinate system is maintained, and no dimensionality reduction technique is applied to the data. By training the model with multiple representative trajectories of LEO - encompassing various inclinations, eccentricities, and altitudes - and testing it with unseen orbital motion patterns, a mean error of around 140 km for the positions and 0.12 km/s for the velocities is achieved. The method offers the advantage of delivering interpretable, accurate, and complex models of orbital motion that can be employed for propagation or as inputs to predictive models for other variables of interest, such as atmospheric drag or the probability of collision in an encounter with a spacecraft or space objects. In conclusion, the work demonstrates the promising potential of using SINDy to discover the equations governing the behaviour of satellites in space. The technique has been successfully applied to uncover PDEs describing the motion of satellites in LEO with high accuracy. The method possesses several advantages over traditional models, including the ability to provide physically interpretable, accurate, and complex models of orbital motion derived from high-entropy datasets. These models can be utilised for propagation or as inputs to predictive models for other variables of interest.

The rapid proliferation of data across diverse fields has accentuated the importance of accurate imputation for missing values. This task is crucial for ensuring data integrity and deriving meaningful insights. In response to this challenge, we present Xputer, a novel imputation tool that adeptly integrates Non-negative Matrix Factorization (NMF) with the predictive strengths of XGBoost. One of Xputer's standout features is its versatility: it supports zero imputation, enables hyperparameter optimization through Optuna, and allows users to define the number of iterations. For enhanced user experience and accessibility, we have equipped Xputer with an intuitive Graphical User Interface (GUI) ensuring ease of handling, even for those less familiar with computational tools. In performance benchmarks, Xputer not only rivals the computational speed of established tools such as IterativeImputer but also often outperforms them in terms of imputation accuracy. Furthermore, Xputer autonomously handles a diverse spectrum of data types, including categorical, continuous, and Boolean, eliminating the need for prior preprocessing. Given its blend of performance, flexibility, and user-friendly design, Xputer emerges as a state-of-the-art solution in the realm of data imputation.

Contrastive, self-supervised learning (SSL) is used to train a model that predicts cancer type from miRNA, mRNA or RPPA expression data. This model, a pretrained FT-Transformer, is shown to outperform XGBoost and CatBoost, standard benchmarks for tabular data, when labelled samples are scarce but the number of unlabelled samples is high. This is despite the fact that the datasets we use have $\mathcal{O}(10^{1})$ classes and $\mathcal{O}(10^{2})-\mathcal{O}(10^{4})$ features. After demonstrating the efficacy of our chosen method of self-supervised pretraining, we investigate SSL for multi-modal models. A late-fusion model is proposed, where each omics is passed through its own sub-network, the outputs of which are averaged and passed to the pretraining or downstream objective function. Multi-modal pretraining is shown to improve predictions from a single omics, and we argue that this is useful for datasets with many unlabelled multi-modal samples, but few labelled unimodal samples. Additionally, we show that pretraining each omics-specific module individually is highly effective. This enables the application of the proposed model in a variety of contexts where a large amount of unlabelled data is available from each omics, but only a few labelled samples.