Machine Learning (ML) is an expressive framework for turning data into computer programs. Across many problem domains -- both in industry and policy settings -- the types of computer programs needed for accurate prediction or optimal control are difficult to write by hand. On the other hand, collecting instances of desired system behavior may be relatively more feasible. This makes ML broadly appealing, but also induces data sensitivities that often manifest as unexpected failure modes during deployment. In this sense, the training data available tend to be imperfect for the task at hand. This thesis explores several data sensitivities of modern machine learning and how to address them. We begin by discussing how to prevent ML from codifying prior human discrimination measured in the training data, where we take a fair representation learning approach. We then discuss the problem of learning from data containing spurious features, which provide predictive fidelity during training but are unreliable upon deployment. Here we observe that insofar as standard training methods tend to learn such features, this propensity can be leveraged to search for partitions of training data that expose this inconsistency, ultimately promoting learning algorithms invariant to spurious features. Finally, we turn our attention to reinforcement learning from data with insufficient coverage over all possible states and actions. To address the coverage issue, we discuss how causal priors can be used to model the single-step dynamics of the setting where data are collected. This enables a new type of data augmentation where observed trajectories are stitched together to produce new but plausible counterfactual trajectories.

Dynamic networks offer an insight of how relational systems evolve. However, modeling these networks efficiently remains a challenge, primarily due to computational constraints, especially as the number of observed events grows. This paper addresses this issue by introducing the Deep Relational Event Additive Model (DREAM) as a solution to the computational challenges presented by modeling non-linear effects in Relational Event Models (REMs). DREAM relies on Neural Additive Models to model non-linear effects, allowing each effect to be captured by an independent neural network. By strategically trading computational complexity for improved memory management and leveraging the computational capabilities of Graphic Processor Units (GPUs), DREAM efficiently captures complex non-linear relationships within data. This approach demonstrates the capability of DREAM in modeling dynamic networks and scaling to larger networks. Comparisons with traditional REM approaches showcase DREAM superior computational efficiency. The model potential is further demonstrated by an examination of the patent citation network, which contains nearly 8 million nodes and 100 million events.

This manuscript enriches the framework of continuous normalizing flows (CNFs) within causal inference, primarily to augment the geometric properties of parametric submodels used in targeted maximum likelihood estimation (TMLE). By introducing an innovative application of CNFs, we construct a refined series of parametric submodels that enable a directed interpolation between the prior distribution $p_0$ and the empirical distribution $p_1$. This proposed methodology serves to optimize the semiparametric efficiency bound in causal inference by orchestrating CNFs to align with Wasserstein gradient flows. Our approach not only endeavors to minimize the mean squared error in the estimation but also imbues the estimators with geometric sophistication, thereby enhancing robustness against misspecification. This robustness is crucial, as it alleviates the dependence on the standard $n^{\frac{1}{4}}$ rate for a doubly-robust perturbation direction in TMLE. By incorporating robust optimization principles and differential geometry into the estimators, the developed geometry-aware CNFs represent a significant advancement in the pursuit of doubly robust causal inference.

There has been significant recent progress in training differentially private (DP) models which achieve accuracy that approaches the best non-private models. These DP models are typically pretrained on large public datasets and then fine-tuned on private downstream datasets that are relatively large and similar in distribution to the pretraining data. However, in many applications including personalization and federated learning, it is crucial to perform well (i) in the few-shot setting, as obtaining large amounts of labeled data may be problematic; and (ii) on datasets from a wide variety of domains for use in various specialist settings. To understand under which conditions few-shot DP can be effective, we perform an exhaustive set of experiments that reveals how the accuracy and vulnerability to attack of few-shot DP image classification models are affected as the number of shots per class, privacy level, model architecture, downstream dataset, and subset of learnable parameters in the model vary. We show that to achieve DP accuracy on par with non-private models, the shots per class must be increased as the privacy level increases. We also show that learning parameter-efficient FiLM adapters under DP is competitive with learning just the final classifier layer or learning all of the network parameters. Finally, we evaluate DP federated learning systems and establish state-of-the-art performance on the challenging FLAIR benchmark.

Livestock feeding behavior is an influential research area for those involved in animal husbandry and agriculture. In recent years, there has been a growing interest in automated systems for monitoring the behavior of ruminants. Despite the developments accomplished in the last decade, there is still much to do and learn about the methods for measuring and analyzing livestock feeding behavior. Automated monitoring systems mainly use motion, acoustic, and image sensors to collect animal behavioral data. The performance evaluation of existing methods is a complex task and direct comparisons between studies are difficult. Several factors prevent a direct comparison, starting from the diversity of data and performance metrics used in the experiments. To the best of our knowledge, this work represents the first tutorial-style review on the analysis of the feeding behavior of ruminants, emphasizing the relationship between sensing methodologies, signal processing and computational intelligence methods. It assesses the main sensing methodologies (i.e. based on movement, sound, images/videos and pressure) and the main techniques to measure and analyze the signals associated with feeding behavior, evaluating their use in different settings and situations. It also highlights the potentiality of automated monitoring systems to provide valuable information that improves our understanding of livestock feeding behavior. The relevance of these systems is increasingly important due to their impact on production systems and research. Finally, the paper closes by discussing future challenges and opportunities in livestock feeding behavior monitoring.

End-to-end Speech Translation (ST) aims to convert speech into target text within a unified model. The inherent differences between speech and text modalities often impede effective cross-modal and cross-lingual transfer. Existing methods typically employ hard alignment (H-Align) of individual speech and text segments, which can degrade textual representations. To address this, we introduce Soft Alignment (S-Align), using adversarial training to align the representation spaces of both modalities. S-Align creates a modality-invariant space while preserving individual modality quality. Experiments on three languages from the MuST-C dataset show S-Align outperforms H-Align across multiple tasks and offers translation capabilities on par with specialized translation models.

Electromyograms (EMG)-based hand gesture recognition systems are a promising technology for human/machine interfaces. However, one of their main limitations is the long calibration time that is typically required to handle new users. The paper discusses and analyses the challenge of cross-subject generalization thanks to an original dataset containing the EMG signals of 14 human subjects during hand gestures. The experimental results show that, though an accurate generalization based on pooling multiple subjects is hardly achievable, it is possible to improve the cross-subject estimation by identifying a robust low-dimensional subspace for multiple subjects and aligning it to a target subject. A visualization of the subspace enables us to provide insights for the improvement of cross-subject generalization with EMG signals.

The weather forecasting system is important for science and society, and significant achievements have been made in applying artificial intelligence (AI) to medium-range weather forecasting. However, existing AI-based weather forecasting models still rely on analysis or reanalysis products from the traditional numerical weather prediction (NWP) systems as initial conditions for making predictions, preventing them from being fully independent systems. As a crucial component of an end-to-end global weather forecasting system, data assimilation is vital in generating initial states for forecasting. In this paper, we present an AI-based data assimilation model, i.e., Adas, for global weather variables, which learns to generate the analysis from the background and sparse observations. Different from existing assimilation methods, Adas employs the gated convolution module to handle sparse observations and the gated cross-attention module for capturing the interactions between observations and background efficiently, which are guided by the confidence matrix to represent the availability and quality of observations. Then, we combine Adas with the advanced AI-based weather forecasting model (i.e., FengWu) and construct the first end-to-end AI-based global weather forecasting system: FengWu-Adas. Experiments demonstrate that Adas can assimilate the simulated global observations with the AI-generated background through a one-year simulation and generate high-quality analysis stably in a cyclic manner. Based on the generated analysis, FengWu-Adas exhibits skillful performance and outperforms the Integrated Forecasting System (IFS) in weather forecasting over seven days.

MineObserver 2.0 is an AI framework that uses Computer Vision and Natural Language Processing for assessing the accuracy of learner-generated descriptions of Minecraft images that include some scientifically relevant content. The system automatically assesses the accuracy of participant observations, written in natural language, made during science learning activities that take place in Minecraft. We demonstrate our system working in real-time and describe a teacher support dashboard to showcase observations, both of which advance our previous work. We present the results of a study showing that MineObserver 2.0 improves over its predecessor both in perceived accuracy of the system's generated descriptions as well as in usefulness of the system's feedback. In future work we intend improve system-generated descriptions, give teachers more control and upgrade the system to perform continuous learning to more effectively and rapidly respond to novel observations made by learners.

Logical reasoning is central to complex human activities, such as thinking, debating, and planning; it is also a central component of many AI systems as well. In this paper, we investigate the extent to which encoder-only transformer language models (LMs) can reason according to logical rules. We ask whether those LMs can deduce theorems in propositional calculus and first-order logic; if their relative success in these problems reflects general logical capabilities; and which layers contribute the most to the task. First, we show for several encoder-only LMs that they can be trained, to a reasonable degree, to determine logical validity on various datasets. Next, by cross-probing fine-tuned models on these datasets, we show that LMs have difficulty in transferring their putative logical reasoning ability, which suggests that they may have learned dataset-specific features, instead of a general capability. Finally, we conduct a layerwise probing experiment, which shows that the hypothesis classification task is mostly solved through higher layers.