An open problem in artificial intelligence is how systems can flexibly learn discrete abstractions that are useful for solving inherently continuous problems. Previous work in computational neuroscience has considered this functional integration of discrete and continuous variables during decision-making under the formalism of active inference (Parr, Friston & de Vries, 2017; Parr & Friston, 2018). However, their focus is on the expressive physical implementation of categorical decisions and the hierarchical mixed generative model is assumed to be known. As a consequence, it is unclear how this framework might be extended to learning. We therefore present a novel hierarchical hybrid active inference agent in which a high-level discrete active inference planner sits above a low-level continuous active inference controller. We make use of recent work in recurrent switching linear dynamical systems (rSLDS) which implement end-to-end learning of meaningful discrete representations via the piecewise linear decomposition of complex continuous dynamics (Linderman et al., 2016). The representations learned by the rSLDS inform the structure of the hybrid decision-making agent and allow us to (1) specify temporally-abstracted sub-goals in a method reminiscent of the options framework, (2) lift the exploration into discrete space allowing us to exploit information-theoretic exploration bonuses and (3) `cache' the approximate solutions to low-level problems in the discrete planner. We apply our model to the sparse Continuous Mountain Car task, demonstrating fast system identification via enhanced exploration and successful planning through the delineation of abstract sub-goals.

Given the growth in the variety and precision of astronomical datasets of interest for cosmology, the best cosmological constraints are invariably obtained by combining data from different experiments. At the likelihood level, one complication in doing so is the need to marginalise over large-dimensional parameter models describing the data of each experiment. These include both the relatively small number of cosmological parameters of interest and a large number of "nuisance" parameters. Sampling over the joint parameter space for multiple experiments can thus become a very computationally expensive operation. This can be significantly simplified if one could sample directly from the marginal cosmological posterior distribution of preceding experiments, depending only on the common set of cosmological parameters. In this paper, we show that this can be achieved by emulating marginal posterior distributions via normalising flows. The resulting trained normalising flow models can be used to efficiently combine cosmological constraints from independent datasets without increasing the dimensionality of the parameter space under study. We show that the method is able to accurately describe the posterior distribution of real cosmological datasets, as well as the joint distribution of different datasets, even when significant tension exists between experiments. The resulting joint constraints can be obtained in a fraction of the time it would take to combine the same datasets at the level of their likelihoods. We construct normalising flow models for a set of public cosmological datasets of general interests and make them available, together with the software used to train them, and to exploit them in cosmological parameter inference.

For statistical modeling wherein the data regime is unfavorable in terms of dimensionality relative to the sample size, finding hidden sparsity in the ground truth can be critical in formulating an accurate statistical model. The so-called "l0 norm" which counts the number of non-zero components in a vector, is a strong reliable mechanism of enforcing sparsity when incorporated into an optimization problem. However, in big data settings wherein noisy estimates of the gradient must be evaluated out of computational necessity, the literature is scant on methods that reliably converge. In this paper we present an approach towards solving expectation objective optimization problems with cardinality constraints. We prove convergence of the underlying stochastic process, and demonstrate the performance on two Machine Learning problems.

Infectious diseases pose significant human and economic burdens. Accurately forecasting disease incidence can enable public health agencies to respond effectively to existing or emerging diseases. Despite progress in the field, developing accurate forecasting models remains a significant challenge. This thesis proposes two methodological frameworks using neural networks (NNs) with associated uncertainty estimates - a critical component limiting the application of NNs to epidemic forecasting thus far. We develop our frameworks by forecasting influenza-like illness (ILI) in the United States. Our first proposed method uses Web search activity data in conjunction with historical ILI rates as observations for training NN architectures. Our models incorporate Bayesian layers to produce uncertainty intervals, positioning themselves as legitimate alternatives to more conventional approaches. The best performing architecture: iterative recurrent neural network (IRNN), reduces mean absolute error by 10.3% and improves Skill by 17.1% on average in forecasting tasks across four flu seasons compared to the state-of-the-art. We build on this method by introducing IRNNs, an architecture which changes the sampling procedure in the IRNN to improve the uncertainty estimation. Our second framework uses neural ordinary differential equations to bridge the gap between mechanistic compartmental models and NNs; benefiting from the physical constraints that compartmental models provide. We evaluate eight neural ODE models utilising a mixture of ILI rates and Web search activity data to provide forecasts. These are compared with the IRNN and IRNN0 - the IRNN using only ILI rates. Models trained without Web search activity data outperform the IRNN0 by 16% in terms of Skill. Future work should focus on more effectively using neural ODEs with Web search data to compete with the best performing IRNN.

Dataset distillation (DD) is an increasingly important technique that focuses on constructing a synthetic dataset capable of capturing the core information in training data to achieve comparable performance in models trained on the latter. While DD has a wide range of applications, the theory supporting it is less well evolved. New methods of DD are compared on a common set of benchmarks, rather than oriented towards any particular learning task. In this work, we present a formal model of DD, arguing that a precise characterization of the underlying optimization problem must specify the inference task associated with the application of interest. Without this task-specific focus, the DD problem is under-specified, and the selection of a DD algorithm for a particular task is merely heuristic. Our formalization reveals novel applications of DD across different modeling environments. We analyze existing DD methods through this broader lens, highlighting their strengths and limitations in terms of accuracy and faithfulness to optimal DD operation. Finally, we present numerical results for two case studies important in contemporary settings. Firstly, we address a critical challenge in medical data analysis: merging the knowledge from different datasets composed of intersecting, but not identical, sets of features, in order to construct a larger dataset in what is usually a small sample setting. Secondly, we consider out-of-distribution error across boundary conditions for physics-informed neural networks (PINNs), showing the potential for DD to provide more physically faithful data. By establishing this general formulation of DD, we aim to establish a new research paradigm by which DD can be understood and from which new DD techniques can arise.

The majority of language model training builds on imitation learning. It covers pretraining, supervised fine-tuning, and affects the starting conditions for reinforcement learning from human feedback (RLHF). The simplicity and scalability of maximum likelihood estimation (MLE) for next token prediction led to its role as predominant paradigm. However, the broader field of imitation learning can more effectively utilize the sequential structure underlying autoregressive generation. We focus on investigating the inverse reinforcement learning (IRL) perspective to imitation, extracting rewards and directly optimizing sequences instead of individual token likelihoods and evaluate its benefits for fine-tuning large language models. We provide a new angle, reformulating inverse soft-Q-learning as a temporal difference regularized extension of MLE. This creates a principled connection between MLE and IRL and allows trading off added complexity with increased performance and diversity of generations in the supervised fine-tuning (SFT) setting. We find clear advantages for IRL-based imitation, in particular for retaining diversity while maximizing task performance, rendering IRL a strong alternative on fixed SFT datasets even without online data generation. Our analysis of IRL-extracted reward functions further indicates benefits for more robust reward functions via tighter integration of supervised and preference-based LLM post-training.

Training social event detection models through federated learning (FedSED) aims to improve participants' performance on the task. However, existing federated learning paradigms are inadequate for achieving FedSED's objective and exhibit limitations in handling the inherent heterogeneity in social data. This paper proposes a personalized federated learning framework with a dual aggregation mechanism for social event detection, namely DAMe. We present a novel local aggregation strategy utilizing Bayesian optimization to incorporate global knowledge while retaining local characteristics. Moreover, we introduce a global aggregation strategy to provide clients with maximum external knowledge of their preferences. In addition, we incorporate a global-local event-centric constraint to prevent local overfitting and ``client-drift''. Experiments within a realistic simulation of a natural federated setting, utilizing six social event datasets spanning six languages and two social media platforms, along with an ablation study, have demonstrated the effectiveness of the proposed framework. Further robustness analyses have shown that DAMe is resistant to injection attacks.

Localization using time-difference of arrival (TDOA) has myriad applications, e.g., in passive surveillance systems and marine mammal research. In this paper, we present a Bayesian estimation method that can localize an unknown number of static sources in 3-D based on TDOA measurements. The proposed localization algorithm based on particle flow (PFL) can overcome the challenges related to the highly nonlinear TDOA measurement model, the data association (DA) uncertainty, and the uncertainty in the number of sources to be localized. Different PFL strategies are compared within a unified belief propagation (BP) framework in a challenging multisensor source localization problem. In particular, we consider PFL-based approximation of beliefs based on one or multiple Gaussian kernels with parameters computed using deterministic and stochastic flow processes. Our numerical results demonstrate that the proposed method can correctly determine the number of sources and provide accurate location estimates. The stochastic flow demonstrates greater accuracy compared to the deterministic flow when using the same number of particles.

We present Flow Matching for Reaction Coordinates (FMRC), a novel deep learning algorithm designed to identify optimal reaction coordinates (RC) in biomolecular reversible dynamics. FMRC is based on the mathematical principles of lumpability and decomposability, which we reformulate into a conditional probability framework for efficient data-driven optimization using deep generative models. While FMRC does not explicitly learn the well-established transfer operator or its eigenfunctions, it can effectively encode the dynamics of leading eigenfunctions of the system transfer operator into its low-dimensional RC space. We further quantitatively compare its performance with several state-of-the-art algorithms by evaluating the quality of Markov State Models (MSM) constructed in their respective RC spaces, demonstrating the superiority of FMRC in three increasingly complex biomolecular systems. Finally, we discuss its potential applications in downstream applications such as enhanced sampling methods and MSM construction.

A natural strategy for continual learning is to weigh a Bayesian ensemble of fixed functions. This suggests that if a (single) neural network could be interpreted as an ensemble, one could design effective algorithms that learn without forgetting. To realize this possibility, we observe that a neural network classifier with N parameters can be interpreted as a weighted ensemble of N classifiers, and that in the lazy regime limit these classifiers are fixed throughout learning. We term these classifiers the neural tangent experts and show they output valid probability distributions over the labels. We then derive the likelihood and posterior probability of each expert given past data. Surprisingly, we learn that the posterior updates for these experts are equivalent to a scaled and projected form of stochastic gradient descent (SGD) over the network weights. Away from the lazy regime, networks can be seen as ensembles of adaptive experts which improve over time. These results offer a new interpretation of neural networks as Bayesian ensembles of experts, providing a principled framework for understanding and mitigating catastrophic forgetting in continual learning settings.