Predicting creep strain is critical for ensuring the durability of adhesively bonded joints. Conventional machine learning methods can provide predictions based on available data, thus reducing the need for time-consuming physical tests. However, these methods often struggle with data scarcity and structural variability, particularly for specific types of joints. This study proposes a deep transfer learning (TL) framework designed for tabular data to address these challenges. The framework was developed using 482 experimental data points from bonded metallic joints for pre-training, while 88 data points from bonded polyethylene joints and 108 data points from bonded carbon fibre-reinforced polymer-steel joints were used for fine-tuning the final TL models. Natural language processing (NLP) concepts were adapted to the tabular creep datasets by treating input features with padding and masking to align heterogeneous feature spaces. These strategies enable the TL framework to transfer knowledge learned despite differences in available input variables. This enhances the robustness and generalisability of transfer learning for sparse, structurally variable tabular engineering data. Compared to the benchmark multilayer perceptron (MLP) models, both the TL and NLP-assisted TL models demonstrated superior performance. Leveraging precious learning experience from the large pre-training dataset, the TL model achieved coefficient of determination (R 2 ) values of 0.89 for the training set and 0.91 for the test set, whereas the MLP model yielded 0.83 (training) and 0.77 (test), indicating potential overfitting. Similar improvements were observed for the NLP-assisted TL model, which effectively managed input feature diversity across joint types. This research underscores the strengths and limitations of the proposed TL and NLP-assisted TL frameworks in overcoming tabular data scarcity and variability in engineering joint studies, offering valuable insights for safety-critical predictions.

Existing infrared small target detection (IRSTD) methods typically assume that training and testing data share the same distribution. However, this assumption often fails in real-world applications due to environmental and sensor-induced variations, resulting in significant performance degradation caused by domain shifts. Besides, the inherently low signal-to-clutter ratio of targets in infrared images further impedes the extraction of underlying target information, increasing the risk of overfitting to domain-specific patterns. This severely constrains the generalizability of knowledge learned from source domains, particularly when training is confined to a single source domain due to the high cost of data annotation. To solve this problem, we propose hierarchical structure dependency whitening (HSDW) for single-domain generalized IRSTD. Specifically, we characterize domain discrepancies in infrared images as differences in structural information. Building upon this point, we employ feature whitening to mitigate the dependency on domain-specific structure information, whose distribution is diversely simulated by a dual-branch nonlinear transformation module. Moreover, we adopt a hierarchical suppression mechanism to alleviate the structural biases across multiple decoding stages, thereby facilitating more generalized target understanding across domains. Extensive experiments on three public IRSTD datasets demonstrate that our method achieves state-of-the-art performance.

Structural variant (SV) detection is essential for genomic studies, and long-read sequencing technologies have advanced our capacity to detect SVs directly from read or de novo assembly, also known as read-based and assembly-based strategy. However, to date, no independent studies have compared and benchmarked the two strategies. Here, on the basis of SVs detected by 20 read-based and eight assembly-based detection pipelines from six datasets of HG002 genome, we investigated the factors that influence the two strategies and assessed their performance with well-curated SVs. We found that up to 80% of the SVs could be detected by both strategies among different long-read datasets, whereas variant type, size, and breakpoint detected by read-based strategy were greatly affected by aligners. For the high-confident insertions and deletions at non-tandem repeat regions, a remarkable subset of them (82% in assembly-based calls and 93% in read-based calls), accounting for around 4000 SVs, could be captured by both reads and assemblies. However, discordance between two strategies was largely caused by complex SVs and inversions, which resulted from inconsistent alignment of reads and assemblies at these loci. Finally, benchmarking with SVs at medically relevant genes, the recall of read-based strategy reached 77% on 5X coverage data, whereas assembly-based strategy required 20X coverage data to achieve similar performance. Therefore, integrating SVs from read and assembly is suggested for general-purpose detection because of inconsistently detected complex SVs and inversions, whereas assembly-based strategy is optional for applications with limited resources.

We study two-layer neural networks in the mean field limit, where the number of neurons tends to infinity. In this regime, the optimization over the neuron parameters becomes the optimization over the probability measures, and by adding an entropic regularizer, the minimizer of the problem is identified as a fixed point. We propose a novel training algorithm named entropic fictitious play, inspired by the classical fictitious play in game theory for learning Nash equilibriums, to recover this fixed point, and the algorithm exhibits a two-loop iteration structure. Exponential convergence is proved in this paper and we also verify our theoretical results by simple numerical examples.