The fine-tuning of pre-trained language models (PLMs) has been shown to be effective across various domains. By using domain-specific supervised data, the general-purpose representation derived from PLMs can be transformed into a domain-specific representation. However, these methods often fail to generalize to out-of-domain (OOD) data due to their reliance on non-causal representations, often described as spurious features. Existing methods either make use of adjustments with strong assumptions about lack of hidden common causes, or mitigate the effect of spurious features using multi-domain data. In this work, we investigate how fine-tuned pre-trained language models aid generalizability from single-domain scenarios under mild assumptions, targeting more general and practical real-world scenarios. We show that a robust representation can be derived through a so-called causal front-door adjustment, based on a decomposition assumption, using fine-tuned representations as a source of data augmentation. Comprehensive experiments in both synthetic and real-world settings demonstrate the superior generalizability of the proposed method compared to existing approaches. Our work thus sheds light on the domain generalization problem by introducing links between fine-tuning and causal mechanisms into representation learning.

Background: In the rapidly evolving landscape of veterinary healthcare, integrating machine learning (ML) clinical decision-making tools with electronic health records (EHRs) promises to improve diagnostic accuracy and patient care. However, the seamless integration of ML classifiers into existing EHRs in veterinary medicine is frequently hindered by the rigidity of EHR systems or the limited availability of IT resources. Results: To address this shortcoming, we present Anna, a freely-available software solution that provides ML classifier results for EHR laboratory data in real-time. Anna is a standalone platform developed in Python, designed to host ML classifiers, retrieve patient-specific data from an EHR system, generate classifier results and return these results to the EHR for display. Anna merges results from different diagnostic tests according to user-defined temporal criteria and determines whether the data are sufficient for a given classifier. Because Anna is a stand-alone platform, it does not require substantial modifications to the existing EHR, allowing for easy integration into existing computing infrastructure. To demonstrate Anna's versatility, we implemented three previously published ML classifiers to predict a diagnosis of hypoadrenocorticism, leptospirosis, or a portosystemic shunt in dogs. Conclusion: Anna is an open-source tool designed to improve the accessibility of ML classifiers for the veterinary community. Its flexible architecture supports the integration of classifiers developed in various programming languages and with diverse environment requirements.

In processing multiple time series, accounting for the individual features of each sequence can be challenging. To address this, modern deep learning methods for time series analysis combine a shared (global) model with local layers, specific to each time series, often implemented as learnable embeddings. Ideally, these local embeddings should encode meaningful representations of the unique dynamics of each sequence. However, when these are learned end-to-end as parameters of a forecasting model, they may end up acting as mere sequence identifiers. Shared processing blocks may then become reliant on such identifiers, limiting their transferability to new contexts. In this paper, we address this issue by investigating methods to regularize the learning of local learnable embeddings for time series processing. Specifically, we perform the first extensive empirical study on the subject and show how such regularizations consistently improve performance in widely adopted architectures. Furthermore, we show that methods preventing the co-adaptation of local and global parameters are particularly effective in this context. This hypothesis is validated by comparing several methods preventing the downstream models from relying on sequence identifiers, going as far as completely resetting the embeddings during training. The obtained results provide an important contribution to understanding the interplay between learnable local parameters and shared processing layers: a key challenge in modern time series processing models and a step toward developing effective foundation models for time series.

K-means clustering is a cornerstone of data mining, but its efficiency deteriorates when confronted with massive datasets. To address this limitation, we propose a novel heuristic algorithm that leverages the Variable Neighborhood Search (VNS) metaheuristic to optimize K-means clustering for big data. Our approach is based on the sequential optimization of the partial objective function landscapes obtained by restricting the Minimum Sum-of-Squares Clustering (MSSC) formulation to random samples from the original big dataset. Within each landscape, systematically expanding neighborhoods of the currently best (incumbent) solution are explored by reinitializing all degenerate and a varying number of additional centroids. Extensive and rigorous experimentation on a large number of real-world datasets reveals that by transforming the traditional local search into a global one, our algorithm significantly enhances the accuracy and efficiency of K-means clustering in big data environments, becoming the new state of the art in the field.

In many application domains, the proliferation of sensors and devices is generating vast volumes of data, imposing significant pressure on existing data analysis and data mining techniques. Nevertheless, an increase in data volume does not inherently imply an increase in informational content, as a substantial portion may be redundant or represent noise. This challenge is particularly evident in the deep learning domain, where the utility of additional data is contingent on its informativeness. In the absence of such, larger datasets merely exacerbate the computational cost and complexity of the learning process. To address these challenges, we propose RAZOR, a novel instance selection technique designed to extract a significantly smaller yet sufficiently informative subset from a larger set of instances without compromising the learning process. RAZOR has been specifically engineered to be robust, efficient, and scalable, making it suitable for large-scale datasets. Unlike many techniques in the literature, RAZOR is capable of operating in both supervised and unsupervised settings. Experimental results demonstrate that RAZOR outperforms recent state-of-the-art techniques in terms of both effectiveness and efficiency.

Accurate uncertainty quantification for causal effects is essential for robust decision making in complex systems, but remains challenging in non-parametric settings. One promising framework represents conditional distributions in a reproducing kernel Hilbert space and places Gaussian process priors on them to infer posteriors on causal effects, but requires restrictive nuclear dominant kernels and approximations that lead to unreliable uncertainty estimates. In this work, we introduce a method, IMPspec, that addresses these limitations via a spectral representation of the Hilbert space. We show that posteriors in this model can be obtained explicitly, by extending a result in Hilbert space regression theory. We also learn the spectral representation to optimise posterior calibration. Our method achieves state-of-the-art performance in uncertainty quantification and causal Bayesian optimisation across simulations and a healthcare application.

We present novel reductions from sample compression schemes in multiclass classification, regression, and adversarially robust learning settings to binary sample compression schemes. Assuming we have a compression scheme for binary classes of size $f(d_\mathrm{VC})$, where $d_\mathrm{VC}$ is the VC dimension, then we have the following results: (1) If the binary compression scheme is a majority-vote or a stable compression scheme, then there exists a multiclass compression scheme of size $O(f(d_\mathrm{G}))$, where $d_\mathrm{G}$ is the graph dimension. Moreover, for general binary compression schemes, we obtain a compression of size $O(f(d_\mathrm{G})\log|Y|)$, where $Y$ is the label space. (2) If the binary compression scheme is a majority-vote or a stable compression scheme, then there exists an $\epsilon$-approximate compression scheme for regression over $[0,1]$-valued functions of size $O(f(d_\mathrm{P}))$, where $d_\mathrm{P}$ is the pseudo-dimension. For general binary compression schemes, we obtain a compression of size $O(f(d_\mathrm{P})\log(1/\epsilon))$. These results would have significant implications if the sample compression conjecture, which posits that any binary concept class with a finite VC dimension admits a binary compression scheme of size $O(d_\mathrm{VC})$, is resolved (Littlestone and Warmuth, 1986; Floyd and Warmuth, 1995; Warmuth, 2003). Our results would then extend the proof of the conjecture immediately to other settings. We establish similar results for adversarially robust learning and also provide an example of a concept class that is robustly learnable but has no bounded-size compression scheme, demonstrating that learnability is not equivalent to having a compression scheme independent of the sample size, unlike in binary classification, where compression of size $2^{O(d_\mathrm{VC})}$ is attainable (Moran and Yehudayoff, 2016).

We theoretically investigate the in-context learning capabilities of transformers in the context of learning mixtures of linear regression models. For the case of two mixtures, we demonstrate the existence of transformers that can achieve an accuracy, relative to the oracle predictor, of order $\mathcal{\tilde{O}}((d/n)^{1/4})$ in the low signal-to-noise ratio (SNR) regime and $\mathcal{\tilde{O}}(\sqrt{d/n})$ in the high SNR regime, where $n$ is the length of the prompt, and $d$ is the dimension of the problem. Additionally, we derive in-context excess risk bounds of order $\mathcal{O}(L/\sqrt{B})$, where $B$ denotes the number of (training) prompts, and $L$ represents the number of attention layers. The order of $L$ depends on whether the SNR is low or high. In the high SNR regime, we extend the results to $K$-component mixture models for finite $K$. Extensive simulations also highlight the advantages of transformers for this task, outperforming other baselines such as the Expectation-Maximization algorithm.

Forecasting of armed conflicts is an important area of research that has the potential to save lives and prevent suffering. However, most existing forecasting models provide only point predictions without any individual-level uncertainty estimates. In this paper, we introduce a novel extension to conformal prediction algorithm which we call bin-conditional conformal prediction. This method allows users to obtain individual-level prediction intervals for any arbitrary prediction model while maintaining a specific level of coverage across user-defined ranges of values. We apply the bin-conditional conformal prediction algorithm to forecast fatalities from armed conflict. Our results demonstrate that the method provides well-calibrated uncertainty estimates for the predicted number of fatalities. Compared to standard conformal prediction, the bin-conditional method outperforms offers improved calibration of coverage rates across different values of the outcome, but at the cost of wider prediction intervals.

Sparse Mixture of Experts (SMoE) has become the key to unlocking unparalleled scalability in deep learning. SMoE has the potential to exponentially increase in parameter count while maintaining the efficiency of the model by only activating a small subset of these parameters for a given sample. However, it has been observed that SMoE suffers from unstable training and has difficulty adapting to new distributions, leading to the model's lack of robustness to data contamination. To overcome these limitations, we first establish a connection between the dynamics of the expert representations in SMoEs and gradient descent on a multi-objective optimization problem. Leveraging our framework, we then integrate momentum into SMoE and propose a new family of SMoEs, named MomentumSMoE. We theoretically prove and numerically demonstrate that MomentumSMoE is more stable and robust than SMoE. In particular, we verify the advantages of MomentumSMoE over SMoE on a variety of practical tasks including ImageNet-1K object recognition and WikiText-103 language modeling. We demonstrate the applicability of MomentumSMoE to many types of SMoE models, including those in the Sparse MoE model for vision (V-MoE) and the Generalist Language Model (GLaM). We also show that other advanced momentum-based optimization methods, such as Adam, can be easily incorporated into the MomentumSMoE framework for designing new SMoE models with even better performance, almost negligible additional computation cost, and simple implementations.