With global wind energy capacity ramping up, accurately predicting damage equivalent loads (DELs) and fatigue across wind turbine populations is critical, not only for ensuring the longevity of existing wind farms but also for the design of new farms. However, the estimation of such quantities of interests is hampered by the inherent complexity in modeling critical underlying processes, such as the aerodynamic wake interactions between turbines that increase mechanical stress and reduce useful lifetime. While high-fidelity computational fluid dynamics and aeroelastic models can capture these effects, their computational requirements limits real-world usage. Recently, fast machine learning-based surrogates which emulate more complex simulations have emerged as a promising solution. Yet, most surrogates are task-specific and lack flexibility for varying turbine layouts and types. This study explores the use of graph neural networks (GNNs) to create a robust, generalizable flow and DEL prediction platform. By conceptualizing wind turbine populations as graphs, GNNs effectively capture farm layout-dependent relational data, allowing extrapolation to novel configurations. We train a GNN surrogate on a large database of PyWake simulations of random wind farm layouts to learn basic wake physics, then fine-tune the model on limited data for a specific unseen layout simulated in HAWC2Farm for accurate adapted predictions. This transfer learning approach circumvents data scarcity limitations and leverages fundamental physics knowledge from the source low-resolution data. The proposed platform aims to match simulator accuracy, while enabling efficient adaptation to new higher-fidelity domains, providing a flexible blueprint for wake load forecasting across varying farm configurations.

The disassembly of power batteries poses significant challenges due to their complex sources, diverse types, variations in design and manufacturing processes, and diverse service conditions. Human memory capacity and robot cognitive and understanding capabilities are limited when faced with different dismantling tasks for end-of-life power batteries. Insufficient human-computer interaction capabilities greatly hinder the efficiency of human-robot collaboration (HRC) operations. The existing HRC relies heavily on the experience of operators, while the existing disassembly system fails to update new disassembly strategies in real time when facing new battery varieties. Therefore, this paper proposes an augmented reality-assisted human-robot collaboration (AR-HRC) power battery dismantling system based on transfer learning. It consists of three modules: AR-HRC knowledge modeling, dismantling subgraph similarity assessment, and strategy transfer update. The AR-HRC knowledge modeling module aims to establish an intelligent mapping from tasks to collaborative strategies based on part features. Based on the evaluation of task similarity, the mobility assessment model divides subtasks into similar and dissimilar classes. For similar subtasks, the original dismantling strategy can be applied to the current task. However, for different subtasks, operators can issue instructions to the AR-HRC system through the human-computer interaction function of AR and develop new collaborative strategies based on actual conditions. Finally, a case study of power battery dismantling is conducted, and the results show that compared to traditional pre-programmed assembly, this system can improve dismantling efficiency and reduce cognitive burden.

This manuscript introduces deep learning models that simultaneously describe the dynamics of several yield curves. We aim to learn the dependence structure among the different yield curves induced by the globalization of financial markets and exploit it to produce more accurate forecasts. By combining the self-attention mechanism and nonparametric quantile regression, our model generates both point and interval forecasts of future yields. The architecture is designed to avoid quantile crossing issues affecting multiple quantile regression models. Numerical experiments conducted on two different datasets confirm the effectiveness of our approach. Finally, we explore potential extensions and enhancements by incorporating deep ensemble methods and transfer learning mechanisms.

Traditionally, electricity distribution networks were designed for unidirectional power flow without the need to accommodate generation installed at the point of use. However, with the increase in Distributed Energy Resources and other Low Carbon Technologies, the role of distribution networks is changing. This shift brings challenges, including the need for intensive metering and more frequent reconfiguration to identify threats from voltage and thermal violations. Mitigating action through reconfiguration is informed by State Estimation, which is especially challenging for low voltage distribution networks where the constraints of low observability, non-linear load relationships, and highly unbalanced systems all contribute to the difficulty of producing accurate state estimates. To counter low observability, this paper proposes the application of a novel transfer learning methodology, based upon the concept of conditional online Bayesian transfer, to make forward predictions of bus pseudo-measurements. Day ahead load forecasts at a fully observed point on the network are adjusted using the intraday residuals at other points in the network to provide them with load forecasts without the need for a complete set of forecast models at all substations. These form pseudo-measurements that then inform the state estimates at future time points. This methodology is demonstrated on both a representative IEEE Test network and on an actual GB 11 kV feeder network.

Developing an artificial design agent that mimics human design behaviors through the integration of heuristics is pivotal for various purposes, including advancing design automation, fostering human-AI collaboration, and enhancing design education. However, this endeavor necessitates abundant behavioral data from human designers, posing a challenge due to data scarcity for many design problems. One potential solution lies in transferring learned design knowledge from one problem domain to another. This article aims to gather empirical evidence and computationally evaluate the transferability of design knowledge represented at a high level of abstraction across different design problems. Initially, a design agent grounded in reinforcement learning (RL) is developed to emulate human design behaviors. A data-driven reward mechanism, informed by the Markov chain model, is introduced to reinforce prominent sequential design patterns. Subsequently, the design agent transfers the acquired knowledge from a source task to a target task using a problem-agnostic high-level representation. Through a case study involving two solar system designs, one dataset trains the design agent to mimic human behaviors, while another evaluates the transferability of these learned behaviors to a distinct problem. Results demonstrate that the RL-based agent outperforms a baseline model utilizing the first-order Markov chain model in both the source task without knowledge transfer and the target task with knowledge transfer. However, the model’s performance is comparatively lower in predicting the decisions of low-performing designers, suggesting caution in its application, as it may yield unsatisfactory results when mimicking such behaviors.

Domain adaptation is important in agriculture because agricultural systems have their own individual characteristics. Applying the same treatment practices (e.g., fertilization) to different systems may not have the desired effect due to those characteristics. Domain adaptation is also an inherent aspect of digital twins. In this work, we examine the potential of transfer learning for domain adaptation in pasture digital twins. We use a synthetic dataset of grassland pasture simulations to pretrain and fine-tune machine learning metamodels for nitrogen response rate prediction. We investigate the outcome in locations with diverse climates, and examine the effect on the results of including more weather and agricultural management practices data during the pretraining phase. We find that transfer learning seems promising to make the models adapt to new conditions. Moreover, our experiments show that adding more weather data on the pretraining phase has a small effect on fine-tuned model performance compared to adding more management practices. This is an interesting finding that is worth further investigation in future studies.

Supervised machine learning is an increasingly popular tool for analyzing large political text corpora. The main disadvantage of supervised machine learning is the need for thousands of manually annotated training data points. This issue is particularly important in the social sciences where most new research questions require new training data for a new task tailored to the specific research question. This paper analyses how deep transfer learning can help address this challenge by accumulating “prior knowledge” in language models. Models like BERT can learn statistical language patterns through pre-training (“language knowledge”), and reliance on task-specific data can be reduced by training on universal tasks like natural language inference (NLI; “task knowledge”). We demonstrate the benefits of transfer learning on a wide range of eight tasks. Across these eight tasks, our BERT-NLI model fine-tuned on 100 to 2,500 texts performs on average 10.7 to 18.3 percentage points better than classical models without transfer learning. Our study indicates that BERT-NLI fine-tuned on 500 texts achieves similar performance as classical models trained on around 5,000 texts. Moreover, we show that transfer learning works particularly well on imbalanced data. We conclude by discussing limitations of transfer learning and by outlining new opportunities for political science research.

Drug discovery uses high throughput screening to identify compounds that interact with a molecular target or that alter a phenotype favorably. The cautious selection of molecules used for such a screening is instrumental and is tightly related to the hit rate. In this work, we wondered if cell painting, a general-purpose image-based assay, could be used as an efficient proxy for compound selection, thus increasing the success rate of a specific assay. To this end, we considered cell painting images with 30,000 molecules treatments, and selected compounds that produced a visual effect close to the positive control of an assay, by using the Frechet Inception Distance. We then compared the hit rates of such a preselection with what was actually obtained in real screening campaigns. As a result, cell painting would have permitted a significant increase in the success rate and, even for one of the assays, would have allowed to reach 80% of the hits with 10 times fewer compounds to test. We conclude that images of a cell painting assay can be directly used for compound selection prior to screening, and we provide a simple quantitative approach in order to do so.

The milk production is strongly influenced by the dairy cow welfare related to a good nutrition and the analysis of the digestibility of feeds allows us to evaluate the health status of the animals. Through faeces’ visual examination it is possible to estimate the quality of diet fed in terms of lacking in fibre or too high in non-structural carbohydrates. The study was carried out in 2021, on four dairy farms in central Italy. The purpose of this work is the classification and evaluation of dairy cow faeces using RGB image analysis through an artificial intelligence (AI) (convolutional neural network (CNN)) algorithm. The main features to analyse are pH, colour and consistency. For the latter two RGB imaging was combined with deep learning and AI to reach objectivity in samples’ evaluation. The images have been captured with several smartphones and cameras, under various light conditions, collecting a data set of 441 images. Images acquired by RGB cameras are then analysed through CNN technology that extracts features and data previously standardized by a faecal score index assigned after a visual analysis and based on five classes. The results achieved with different training strategies show a training accuracy of 90% and a validation accuracy of 78% of the model which allow us to identify problems in bovine digestion and to intervene promptly in feed variation. The method used in this study eliminates subjectivity in field analysis and allows future improvement of increasing the data set to strengthen the model.

Recent years have seen a surge in interest in building deep learning-based fully data-driven models for weather prediction. Such deep learning models, if trained on observations can mitigate certain biases in current state-of-the-art weather models, some of which stem from inaccurate representation of subgrid-scale processes. However, these data-driven models, being over-parameterized, require a lot of training data which may not be available from reanalysis (observational data) products. Moreover, an accurate, noise-free, initial condition to start forecasting with a data-driven weather model is not available in realistic scenarios. Finally, deterministic data-driven forecasting models suffer from issues with long-term stability and unphysical climate drift, which makes these data-driven models unsuitable for computing climate statistics. Given these challenges, previous studies have tried to pre-train deep learning-based weather forecasting models on a large amount of imperfect long-term climate model simulations and then re-train them on available observational data. In this article, we propose a convolutional variational autoencoder (VAE)-based stochastic data-driven model that is pre-trained on an imperfect climate model simulation from a two-layer quasi-geostrophic flow and re-trained, using transfer learning, on a small number of noisy observations from a perfect simulation. This re-trained model then performs stochastic forecasting with a noisy initial condition sampled from the perfect simulation. We show that our ensemble-based stochastic data-driven model outperforms a baseline deterministic encoder–decoder-based convolutional model in terms of short-term skills, while remaining stable for long-term climate simulations yielding accurate climatology.

Artificial intelligence has provided many breakthroughs in the field of computer vision. The fully convolutional networks U-Net in particular have provided very promising results in the problem of retrieving rain rates from space-borne observations, a challenge that has persisted over the past few decades. The rain intensity is estimated from the measurement of the brightness temperatures on different microwave channels. However, these channels are slightly different depending on the satellite. In the case where a retrieval model has been developed from a single satellite, it may be advantageous to use domain adaptation methods in order to make this model compatible with all the satellites of the constellation. In this proposed feasibility study, a Cycle Generative Adversarial Nets model is used for adapting one set of brightness temperature channels to another set. Results of a toy experiment show that this method is able to provide qualitatively good precipitation structure but still could be improved in terms of precision.

Current research in yes/no question answering (QA) focuses on transfer learning techniques and transformer-based models. Models trained on large corpora are fine-tuned on tasks similar to yes/no QA, and then the captured knowledge is transferred for solving the yes/no QA task. Most previous studies use existing similar tasks, such as natural language inference or extractive QA, for the fine-tuning step. This paper follows a different perspective, hypothesizing that an artificial yes/no task can transfer useful knowledge for improving the performance of yes/no QA. We introduce three such tasks for this purpose, by adapting three corresponding existing tasks: candidate answer validation, sentiment classification, and lexical simplification. Furthermore, we experimented with three different variations of the BERT model (BERT base, RoBERTa, and ALBERT). The results show that our hypothesis holds true for all artificial tasks, despite the small size of the corresponding datasets that are used for the fine-tuning process, the differences between these tasks, the decisions that we made to adapt the original ones, and the tasks’ simplicity. This gives an alternative perspective on how to deal with the yes/no QA problem, that is more creative, and at the same time more flexible, as it can exploit multiple other existing tasks and corresponding datasets to improve yes/no QA models.

Recent research has reported that standard fine-tuning approaches can be unstable due to being prone to various sources of randomness, including but not limited to weight initialization, training data order, and hardware. Such brittleness can lead to different evaluation results, prediction confidences, and generalization inconsistency of the same models independently fine-tuned under the same experimental setup. Our paper explores this problem in natural language inference, a common task in benchmarking practices, and extends the ongoing research to the multilingual setting. We propose six novel textual entailment and broad-coverage diagnostic datasets for French, German, and Swedish. Our key findings are that the mBERT model demonstrates fine-tuning instability for categories that involve lexical semantics, logic, and predicate-argument structure and struggles to learn monotonicity, negation, numeracy, and symmetry. We also observe that using extra training data only in English can enhance the generalization performance and fine-tuning stability, which we attribute to the cross-lingual transfer capabilities. However, the ratio of particular features in the additional training data might rather hurt the performance for model instances. We are publicly releasing the datasets, hoping to foster the diagnostic investigation of language models (LMs) in a cross-lingual scenario, particularly in terms of benchmarking, which might promote a more holistic understanding of multilingualism in LMs and cross-lingual knowledge transfer.

The choice of a copula model from limited data is a hard but important task. Motivated by the visual patterns that different copula models produce in smoothed density heatmaps, we consider copula model selection as an image recognition problem. We extract image features from heatmaps using the pre-trained AlexNet and present workflows for model selection that combine image features with statistical information. We employ dimension reduction via Principal Component and Linear Discriminant Analyses and use a Support Vector Machine classifier. Simulation studies show that the use of image data improves the accuracy of the copula model selection task, particularly in scenarios where sample sizes and correlations are low. This finding indicates that transfer learning can support statistical procedures of model selection. We demonstrate application of the proposed approach to the joint modelling of weekly returns of the MSCI and RISX indices.

In this paper, we propose a set of robust training methods for deep reinforcement learning to transfer learning acquired in one control task to a set of previously unseen control tasks. We improve generalization in commonly used transfer learning benchmarks by a novel sample elimination technique, early stopping, and maximum entropy adversarial reinforcement learning. To generate robust policies, we use sample elimination during training via a method we call strict clipping. We apply early stopping, a method previously used in supervised learning, to deep reinforcement learning. Subsequently, we introduce maximum entropy adversarial reinforcement learning to increase the domain randomization during training for a better target task performance. Finally, we evaluate the robustness of these methods compared to previous work on simulated robots in target environments where the gravity, the morphology of the robot, and the tangential friction coefficient of the environment are altered.

Vaginitis is a prevalent gynecologic disease that threatens millions of women’s health. Although microscopic examination of vaginal discharge is an effective method to identify vaginal infections, manual analysis of microscopic leucorrhea images is extremely time-consuming and labor-intensive. To automate the detection and identification of visible components in microscopic leucorrhea images for early-stage diagnosis of vaginitis, we propose a novel end-to-end deep learning-based cells detection framework using attention-based detection with transformers (DETR) architecture. The transfer learning was applied to speed up the network convergence while maintaining the lowest annotation cost. To address the issue of detection performance degradation caused by class imbalance, the weighted sampler with on-the-fly data augmentation module was integrated into the detection pipeline. Additionally, the multi-head attention mechanism and the bipartite matching loss system of the DETR model perform well in identifying partially overlapping cells in real-time. With our proposed method, the pipeline achieved a mean average precision (mAP) of 86.00% and the average precision (AP) of epithelium, leukocyte, pyocyte, mildew, and erythrocyte was 96.76, 83.50, 74.20, 89.66, and 88.80%, respectively. The average test time for a microscopic leucorrhea image is approximately 72.3 ms. Currently, this cell detection method represents state-of-the-art performance.

In this study, we investigate the process of generating single-sentence representations for the purpose of Dialogue Act (DA) classification, including several aspects of text pre-processing and input representation which are often overlooked or underreported within the literature, for example, the number of words to keep in the vocabulary or input sequences. We assess each of these with respect to two DA-labelled corpora, using a range of supervised models, which represent those most frequently applied to the task. Additionally, we compare context-free word embedding models with that of transfer learning via pre-trained language models, including several based on the transformer architecture, such as Bidirectional Encoder Representations from Transformers (BERT) and XLNET, which have thus far not been widely explored for the DA classification task. Our findings indicate that these text pre-processing considerations do have a statistically significant effect on classification accuracy. Notably, we found that viable input sequence lengths, and vocabulary sizes, can be much smaller than is typically used in DA classification experiments, yielding no significant improvements beyond certain thresholds. We also show that in some cases the contextual sentence representations generated by language models do not reliably outperform supervised methods. Though BERT, and its derivative models, do represent a significant improvement over supervised approaches, and much of the previous work on DA classification.