Identifying unique polyps in colon capsule endoscopy (CCE) images is a critical yet challenging task for medical personnel due to the large volume of images, the cognitive load it creates for clinicians, and the ambiguity in labeling specific frames. This paper formulates this problem as a multi-instance learning (MIL) task, where a query polyp image is compared with a target bag of images to determine uniqueness. We employ a multi-instance verification (MIV) framework that incorporates attention mechanisms, such as variance-excited multi-head attention (VEMA) and distance-based attention (DBA), to enhance the model's ability to extract meaningful representations. Additionally, we investigate the impact of self-supervised learning using SimCLR to generate robust embeddings. Experimental results on a dataset of 1912 polyps from 754 patients demonstrate that attention mechanisms significantly improve performance, with DBA L1 achieving the highest test accuracy of 86.26% and a test AUC of 0.928 using a ConvNeXt backbone with SimCLR pretraining. This study underscores the potential of MIL and self-supervised learning in advancing automated analysis of Colon Capsule Endoscopy images, with implications for broader medical imaging applications.

Coastline detection in remote sensing imagery is commonly formulated as a pixel-wise segmentation problem, where the final coastline is extracted from a predicted mask through post-processing. This formulation relegates coastline geometry, the primary representation used in coastal change analysis, to a secondary artifact rather than the learning objective. In practice, coastlines are defined by geomorphic proxies such as vegetation lines, dune toes, or cliff edges, rather than an instantaneous land-water boundary often used in pixel-based segmentation approaches. In this work, we revisit coastline extraction from a representation perspective and formulate the task as geometric boundary localization. We use the New Zealand Coastal Change Dataset (NZCCD) and high-resolution aerial imagery from Land Information New Zealand (LINZ) to develop CoastlineVLM-7B, a vision-language model (VLM) built on the GeoChat-7B/LLaVA-1.5 architecture that jointly performs coastline presence detection, proxy-type classification, and coastline grounding. The model directly predicts a coastline as a polyline rather than a dense segmentation mask. We evaluate CoastlineVLM-7B against segmentation baselines under strict one-pixel boundary supervision. Results show that geometry-based metrics are more suitable for assessing coastline localization quality than pixel-overlap metrics such as Intersection over Union (IoU). CoastlineVLM-7B improves global geometric alignment with reference coastlines, reducing Hausdorff distance from 37.74 m to 31.84 m and Earth Mover's Distance from 21.12 m to 17.32 m. These results indicate that output representation is a critical design choice in coastline extraction, and that geometry-oriented learning, combined with the semantic reasoning capabilities of vision-language models, aligns well with how coastlines are defined and evaluated in operational coastal monitoring.

Memory is a critical component in replay-based continual learning (CL). Prior research has largely treated CL memory as a monolithic store of past data, focusing on how to select and store representative past examples. However, this perspective overlooks the higher-level memory architecture that governs the interaction between old and new data. In this work, we identify and characterize a dual-memory system that is inherently present in both online and offline CL settings. This system comprises: a short-term memory, which temporarily buffers recent data for immediate model updates, and a long-term memory, which maintains a carefully curated subset of past experiences for future replay and consolidation. We propose \textit{memory capacity ratio} (MCR), the ratio between short-term memory and long-term memory capacities, to characterize online and offline CL. Based on this framework, we systematically investigate how MCR influences generalization, stability, and plasticity. Across diverse CL settings—class-incremental, task-incremental, and domain-incremental—and multiple data modalities (e.g., image and text classification), we observe that a smaller MCR, characteristic of \textit{online CL}, can yield comparable or even superior performance relative to a larger one, characteristic of \textit{offline CL}, when both are evaluated under equivalent computational and data storage budgets. This advantage holds consistently across several state-of-the-art replay strategies, such as ER, DER, and SCR. Theoretical analysis further reveals that a reduced MCR yields a better trade-off between stability and plasticity by lowering a bound on generalization error when learning from non-stationary data streams with limited memory. These findings offer new insights into the role of memory allocation in continual learning and underscore the underexplored potential of online CL approaches.

Model trees provide an appealing way to perform interpretable machine learning for both classification and regression problems. In contrast to “classic” decision trees with constant values in their leaves, model trees can use linear combinations of predictor variables in their leaf nodes to form predictions, which can help achieve higher accuracy and smaller trees. Typical algorithms for learning model trees from training data work in a greedy fashion, growing the tree in a top-down manner by recursively splitting the data into smaller and smaller subsets. This yields a fast algorithm, but the selected splits are only locally optimal, potentially rendering the tree overly complex and less accurate than a tree whose structure is globally optimal for the training data. In this paper, we empirically investigate the effect of constructing globally optimal model trees for classification and regression. The trees we consider feature linear support vector machines at the leaf nodes and are learned using mixed-integer linear programming (MILP) formulations. We use benchmark datasets to compare them to model trees obtained using greedy and dynamic programming-based algorithms, evaluating both tree size and predictive accuracy. We also compare to classic optimal and greedily grown decision trees, random forests, and support vector machines. Our results show that MILP-based optimal model trees can achieve competitive accuracy with very small trees. We also investigate the effect on the accuracy of replacing axis-parallel splits with multivariate ones, foregoing interpretability while potentially obtaining greater accuracy.

Despite their success, large pretrained vision models remain vulnerable to catastrophic forgetting when adapted to new tasks in class-incremental settings. Parameter-efficient fine-tuning (PEFT) alleviates this by restricting trainable parameters, yet most approaches still rely on cross-entropy (CE) loss, a surrogate for the 0-1 loss, to learn from new data. We revisit this choice and revive the true objective (0-1 loss) through a reinforcement learning perspective. By formulating classification as a one-step Markov Decision Process, we derive an Expected Policy Gradient (EPG) method that directly minimizes misclassification error with a low-variance gradient estimation. Our analysis shows that CE can be interpreted as EPG with an additional sample-weighting mechanism: CE encourages exploration by emphasizing low-confidence samples, while EPG prioritizes high-confidence ones. Building on this insight, we propose adaptive entropy annealing (aEPG), a training strategy that transitions from exploratory (CE-like) to exploitative (EPG-like) learning. aEPG-based methods outperform CE-based methods across diverse benchmarks and with various PEFT modules. More broadly, we evaluate various entropy regularization methods and demonstrate that lower entropy of the output prediction distribution enhances adaptation in pretrained vision models.

Near-infrared (NIR) spectroscopy is a widely used technology in the horticulture industry for non-destructive fruit grading. Partial Least Squares (PLS) regression is the dominant method for producing fruit quality predictions from measured spectra. Alternative deep learning methods have shown promise, but often require large amounts of labelled data to train. This study proposes a semi-supervised method based on Barlow Twins to include unlabelled data in the training process. We adopt the Barlow Twins method by using repeated measurements on the same fruit from different devices as different “views” to encode into the same latent space and combine the encoder network with a regression head for prediction. Our approach demonstrates improved performance over PLS with up to 17% lower RMSE, especially when the labelled data is limited. The Barlow loss function also improves calibration transfer results.

This study aims to predict a patient’s length of stay (LoS) and discharge outcome using parameters collected at the time of hospital admission.

We explore multiple instance verification, a problem setting in which a query instance is verified against a bag of target instances with heterogeneous, unknown relevancy. We show that naive adaptations of attention-based multiple instance learning (MIL) methods and standard verification methods like Siamese neural networks are unsuitable for this setting: directly combining state-of-the-art (SOTA) MIL methods and Siamese networks is shown to be no better, and sometimes significantly worse, than a simple baseline model. Postulating that this may be caused by the failure of the representation of the target bag to incorporate the query instance, we introduce a new pooling approach named "crossattention pooling" (CAP). Under the CAP framework, we propose two novel attention functions to address the challenge of distinguishing between highly similar instances in a target bag. Through empirical studies on three different verification tasks, we demonstrate that CAP outperforms adaptations of SOTA MIL methods and the baseline by substantial margins, in terms of both classification accuracy and the ability to detect key instances. The superior ability to identify key instances is attributed to the new attention functions by ablation studies. We share our code at https://github.com/xxweka/MIV.

Image classification for species identification has applications in areas such as conservation and education. Given New Zealand's geographic isolation and the relatively small number of species present on its islands, there is an opportunity to apply machine learning to enable accurate automatic species identification for Aotearoa, even on mobile devices without Internet access. We present neural network-based image classification models trained to classify organisms present in New Zealand. The data for model development and evaluation, obtained from the crowd-sourcing website iNaturalist, comprises 14,991 species, including 6,216 Animalia, 6,173 Plantae, and 2,407 Fungi species, alongside a small set of observations of Bacteria, Chromista, Protozoa, and Viruses. It contains organisms observed in the natural environment as well as captive and cultivated organisms. The trained models achieve over 76% classification accuracy across all species and produce class probability estimates, calibrated using temperature scaling, that can be used to gauge confidence in their classifications. Input attribution methods can be used to interpret a model's inferences by highlighting its areas of focus on images. The models are available to the public as downloadable model files and as part of both web and mobile applications for species identification that are distributed as open-source software.

As a dominant terrestrial ecosystem, forests play a pivotal role, which is substantially challenged by climate extremes. At the same time, the practice of patient science to investigate and understand different intricate climate-driven phenomena is no longer an option. On the other hand, recent technological advancements enable scientists to simultaneously collect and analyse a large volume of complex data. High-resolution tree stem radius measurements and predictive simulation through machine learning algorithms offer powerful opportunities for understanding these dynamics. However, when these machine learning methods are applied without careful consideration of data quality, model biases, and other critical factors, their potential is often compromised. In this study, we aimed to evaluate four Deep Learning algorithms (namely CNN, LSTM, Transformer, and ETSFormer), using automatically measured and high temporal resolution tree stem radius data. We explore the complexities of handling voluminous and authentic datasets required by these algorithms. Initial experiments show that it is possible to achieve an MAE as small as 0.0026 mm on the full data. Furthermore, our study delves into the temporal resolution of data, demonstrating the feasibility of using reduced datasets without compromising accuracy levels. Our best result showed that a reduction of 97 % in collection events increases the MAE by only 6 % with the LSTM model, demonstrating that resource use optimisation can be achieved by slightly reducing the temporal resolution of data collection with marginal error increase. This also shows that LSTM can effectively capture longer-term and complex dependencies, which indicates promising future work with additional environmental data.

Near infrared (NIR) spectroscopy is widely used as a tool for non-destructive assessment of fruit quality by applying measured spectra to predict quality parameters such as dry matter and soluble solids content using a suitable regression method. With continued advancements in deep learning, there is potential for improved predictive performance when neural network models are applied instead of partial least-squares regression, but choosing a model remains challenging as performance is sensitive to the model's architecture. Taking inspiration from work done in image classification, we propose model selection by assessing relative stability to diffeomorphic transformations, providing a complementary approach to standard validation methods. This is particularly useful when labelled validation data is limited. Our empirical results on several NIR regression problems indicate that the proposed approach is comparable to the use of independent validation sets. In addition to the choice of deep learning architecture, we also consider the selection of the number of components in partial least-squares regression to demonstrate the method's generality.

Recent work [ 14 ] has introduced a method for prior elicitation that utilizes records of expert decisions to infer a prior distribution. While this method provides a promising approach to eliciting expert uncertainty, it has only been demonstrated using tabular data, which may not entirely represent the information used by experts to make decisions. In this paper, we demonstrate how analysts can adopt a deep learning approach to utilize the method proposed in [14 ] with the actual information experts use. We provide an overview of deep learning models that can effectively model expert decision-making to elicit distributions that capture expert uncertainty and present an example examining the risk of colon cancer to show in detail how these models can be used.

We investigate cross-domain few-shot learning under the constraint that fine-tuning of backbones (i.e., feature extractors) is impossible or infeasible – a scenario that is increasingly common in practical use cases. Handling the low-quality and static embeddings produced by frozen, "black-box" backbones leads to a problem representation of few-shot classification as a series of multiple instance verification (MIV) tasks. Inspired by this representation, we introduce a novel approach to few-shot domain adaptation, named the "MIV-head", akin to a classification head that is agnostic to any pretrained backbone and computationally efficient. The core components designed for the MIV-head, when trained on few-shot data from a target domain, collectively yield strong performance on test data from that domain. Importantly, it does so without fine-tuning the backbone, and within the "meta-testing" phase. Experimenting under various settings and on an extension of the Meta-dataset benchmark for cross-domain few-shot image classification, using representative off-the-shelf convolutional neural network and vision transformer backbones pretrained on ImageNet1K, we show that the MIV-head achieves highly competitive accuracy when compared to state-of-the-art "adapter" (or partially fine-tuning) methods applied to the same backbones, while incurring substantially lower adaptation cost. We also find well-known "classification head" approaches lag far behind in terms of accuracy. Ablation study empirically justifies the core components of our approach. We share our code at https://github.com/xxweka/MIV-head.

This article introduces a new method for eliciting prior distributions from experts. The method models an expert decision-making process to infer a prior probability distribution for a rare event A. More specifically, assuming there exists a decision-making process closely related to A which forms a decision Y, where a history of decisions have been collected. By modelling the data observed to make the historic decisions, using a Bayesian model, an analyst can infer a distribution for the parameters of the random variable Y. This distribution can be used to approximate the prior distribution for the parameters of the random variable for event A. This method is novel in the field of prior elicitation and has the potential of improving upon current methods by using real-life decision-making processes, that can carry real-life consequences, and, because it does not require an expert to have statistical knowledge. Future decision making can be improved upon using this method, as it highlights variables that are impacting the decision making process. An application for eliciting a prior distribution of recidivism, for an individual, is used to explain this method further.

In the field of aerial and satellite remote sensing, the widespread adoption of deep learning brings new possibilities. Current approaches, however, often overlook the unique characteristics of aerial data. This study introduces a methodology that capitalizes on distinctive features, leveraging additional annotations for enhanced neural network training. Despite modest gains in classification accuracy, the synergy of enhanced explainability, automated segmentation, and targeted classification demonstrates nuanced improvements. Preliminary results showcase potential applications in land cover mapping. This work can be extented towards reducing dependency on labor-intensive human annotations through an iterative annotation and training loop.

Instead of training deep neural networks offline with a large static dataset, continual learning (CL) considers a new learning paradigm, which continually trains the deep networks from a non-stationary data stream on the fly. Despite the recent progress, continual learning remains an open challenge. Many CL techniques still require offline training of large batches of data chunks (i.e., tasks) over multiple epochs. Conventional wisdom holds that online continual learning, which assumes single-pass data, is strictly harder than offline continual learning, due to the combined challenges of catastrophic forgetting and underfitting within a single training epoch. Here, we challenge this assumption by empirically demonstrating that online CL can match or exceed the performance of its offline counterpart given equivalent memory and computational resources. This finding is further verified across different CL approaches and benchmarks. To better understand these counterintuitive experimental findings, we design a framework to unify and interpolate between online and offline CL and provide a theoretical analysis showing that online CL can yield a tighter generalization bound than offline CL.

As a dominant terrestrial ecosystem, forests play a pivotal role, which is substantially challenged by climate extremes. At the same time, the practice of patient science to investigate and understand different intricate climate-driven phenomena is no longer an option. On the other hand, recent technological advancements enable scientists to simultaneously collect and analyse a large volume of complex data. High-resolution tree stem radius measurements and predictive simulation through machine learning algorithms are employed randomly. However, these opportunities require careful utilisation to realise their full potential. In this study, we aimed to evaluate three state-of-the-art machine learning algorithms through automatically measured and high temporal resolution tree stem radius data. We explore the complexities of handling voluminous and authentic datasets required by these algorithms. Furthermore, our study delves into the temporal resolution of data, demonstrating the feasibility of using reduced datasets without compromising accuracy levels. Our results showed that a reduction of 97% in collection events increases the MAE by only 6%, demonstrating that resource use optimisation can be achieved by slightly reducing the temporal resolution of data collection with marginal error.

State-of-the-art cross-domain few-shot learning methods for image classification apply knowledge transfer by fine-tuning deep feature extractors obtained from source domains on the small labelled dataset available for the target domain, generally in conjunction with a simple centroid-based classification head. Semi-supervised learning during the meta-test phase is an obvious approach to incorporating unlabelled data into cross-domain few-shot learning, but semi-supervised methods designed for larger sets of labelled data than those available in few-shot learning appear to easily go astray when applied in this setting. We propose an efficient semi-supervised learning method that applies self-training to the classification head only and show that it can yield very consistent improvements in average performance in the Meta-Dataset benchmark for cross-domain few-shot learning when applied with contemporary methods utilising centroid-based classification.