As a complex decision-making process, roadnetwork simplification involves stroke recognition, mesh density relatively preserving, and network structure abstracting. Such a multi-factor decision and scaling operation traditionally applied rule-based methods. The construction and adjusting of these rules contain many human-set parameters and conditions, which makes generalized results closely related to the cartographer’sexperience and habits. On the other hand, the existing methods tend to consider individual structures, for example,strokes, meshes, graph networks, etc.,separately in differentalgorithms lacking a solution that bringsthe advantages of these pattern structure handlingstogether. Aiming at the above problems, this study designs a simplification method using the Mesh-Line Structure Unit (MLSU) to simultaneously account for polyline and polygon properties. A graph-based deep learning network is built to use data-driven ideas to realize road selection decisions. The MLSU model can extract22 kinds of polyline features, 5 kinds of polygon features, and 3 interactivefeatures. In order to make generalization decisions,a model based on graph convolutional network is constructed,and the network model is trained with real data from partial areas in the southern United States, thus realizing automatic generalization of the road network. The experimental results show that the proposed method effectively realizes the automatic generalization of road data, and the simplified results have better performance in terms of visual representation, quantity maintenance, and average connectivity compared with other methods. This study also demonstrates the advantages and potential of using graph deep learning techniquesfor map generalization problems.

272,700 two-alternative forced choice responses in a simple numerical task modeled after Tenenbaum (1999, 2000), collected from 606 Amazon Mechanical Turk workers. Subjects were shown sets of numbers length 1 to 4 from the range 1 to 100 (e.g. {12, 16}), and asked what other numbers were likely to belong to that set (e.g. 1, 5, 2, 98). Their generalization patterns reflect both rule-like (e.g. “even numbers,” “powers of two”) and distance-based (e.g. numbers near 50) generalization. This data set is available for further analysis of these simple and intuitive inferences, developing of hands-on modeling instruction, and attempts to understand how probability and rules interact in human cognition.

Topographic databases normally contain areas of different land cover classes, commonly defining a planar partition, that is, gaps and overlaps are not allowed. When reducing the scale of such a database, some areas become too small for representation and need to be aggregated. This unintentionally but unavoidably results in changes of classes. In this article we present an optimisation method for the aggregation problem. This method aims to minimise changes of classes and to create compact shapes, subject to hard constraints ensuring aggregates of sufficient size for the target scale. To quantify class changes we apply a semantic distance measure. We give a graph theoretical problem formulation and prove that the problem is NP-hard, meaning that we cannot hope to find an efficient algorithm. Instead, we present a solution by mixed-integer programming that can be used to optimally solve small instances with existing optimisation software. In order to process large datasets, we introduce specialised heuristics that allow certain variables to be eliminated in advance and a problem instance to be decomposed into independent sub-instances. We tested our method for a dataset of the official German topographic database ATKIS with input scale 1:50,000 and output scale 1:250,000. For small instances, we compare results of this approach with optimal solutions that were obtained without heuristics. We compare results for large instances with those of an existing iterative algorithm and an alternative optimisation approach by simulated annealing. These tests allow us to conclude that, with the defined heuristics, our optimisation method yields high-quality results for large datasets in modest time.

In sports, the role of backswing is considered critical for generating a good shot, even though it plays no direct role in hitting the ball. We recently demonstrated the scientific basis of this phenomenon by showing that immediate past movement affects the learning and recall of motor memories. This effect occurred regardless of whether the past contextual movement was performed actively, passively, or shown visually. In force field studies, it has been shown that motor memories generalize locally and that the level of compensation decays as a function of movement angle away from the trained movement. Here we examine if the contextual effect of past movement exhibits similar patterns of generalization and whether it can explain behavior seen in interference studies. Using a single force-field learning task, the directional tuning curves of both the prior contextual movement and the subsequent force field adaptive movements were measured. The adaptation movement direction showed strong ...

Source data selection for out-of-domain generalization

Successful language acquisition relies on generalisation, yet many 'sensible' generalisations are actually ungrammatical (eg 'John carried me teddy.'). This grant explores how language learners balance generalisation and exception learning using the Artificial Language Learning (ALL) methodology, ie experiments where participants learn and are tested on novel experimenter-designed languages. Earlier research (Wonnacott et al. 2008) had used only adults - an important limitation given evidence for maturational differences in language learning (Newport, 1990). This grant therefore consists of a series of ALL experiments conducted with both child and adult participants, designed to address the following questions: (i) Do children, (like adults in previous studies) use distributional statistics eg word and construction frequency to determine which words should generalise/are exceptions? (ii) How do learners weigh such information against other sources of information such as semantics (eg if words with similar meanings tend to behave similarly). (iii) Do these processes differ across adults and children? (iv) Are there any factors that predict the extent of generalisation/exception learning for individual learners (eg working memory)? The long-term goal is to shed light on why language learning is generally more successful when it begins in childhood and the loci of individual differences in learning.

The automatic adaptive signature generalization (AASG) algorithm overcomes many of the limitations associated with classification of multitemporal imagery. By locating stable sites between two images and using them to adapt class spectral signatures from a high-quality reference classification to a new image, AASG mitigates the impacts of radiometric and phenological differences between images and ensures that class definitions remain consistent between the two classifications. Here, I provide source code (in the R programming environment), as well as a comprehensive user guide, for the AASG algorithm. See Dannenberg, Hakkenberg and Song (2016) for details of the algorithm.

This Zenodo repository contains 100 copies of the model BERT fine-tuned on the MNLI dataset, created for the paper "BERTs of a feather do not generalize together: Large variability in generalization across models with similar test set performance." Please see the project GitHub page for more details about using these models and how to cite any such usage: https://github.com/tommccoy1/hans/tree/master/berts_of_a_feather

Full results of simulations detailed in thesis, Belcher, D 2025, 'Towards an understanding of generalisation in deep learning: an analysis of the transformation of information in convolutional neural networks', Master of Philosophy, University of Southampton, Southampton, UK. This dataset is the results of the simulations detailed in the above thesis. All results are in jsonlines format. No specialist software is required to read this data, any software for parsing json or jsonlines data is sufficient.

### Research Objective

The primary objective of this research is to enhance the generalization of physics-informed machine learning (PIML) models by integrating them with neural oscillators. The goal is to improve the accuracy of these models in predicting solutions to partial differential equations (PDEs) beyond the training domain.

### Type of Research

This research is applied and experimental. It focuses on developing and validating a new methodological approach to enhance the generalization capabilities of PIML models through a series of numerical experiments on various nonlinear and high order PDEs.

### Method of Data Collection

The data used for validating numerical experiments are closed form analytic solution and physics-informed method is utilized to simulate the dataset. Both are explicitly mention in the python notebooks. The experiments are conducted on time-dependent nonlinear PDEs, including the viscous Burgers equation, Allen-Cahn equation, nonlinear Schrödinger equation, Euler-Bernoulli beam equation, and a 2D Kovasznay flow.

### Type of Data/codes

1. All implementation are done using jupyter notebooks (.ipynb) or .py.

2. .mat files are analytical solution generated using PINN simulation.

3. (.jpeg), (.pdf) are figures which are used in the main manuscript.

Stacked generalization of random forest and decision tree techniques for library data visualization The huge amount of library data stored in our modern research and statistic centers of organizations is springing up on daily bases These databases grow exponentially in size with respect to time it becomes exceptionally difficult to easily understand the behavior and interpret data with the relationships that exist between attributes This exponential growth of data poses new organizational c

"graph_data.xlsx" is an excel spreadsheet containing the graph data. There are two sheets, "Nonlinearities" which contains the data in Fig 2a, and "Dispersion" containing the data in Fig 2b. In each sheet, the first column is the X axis and further columns are the Y values.

No description was included in this Dataset collected from the OSF

This is the data used to reproduce the results from "Evaluating natural language processing models with generalization metrics that do not need access to any training or testing data".

This dataset is generated the KAT data center in Paderborn University with the sampling rate of 64 KHz (Lessmeier et al. 2016). The damages were generated using both artificial and natural ways. More specifically, an electric discharge machine (EDM), a drilling, and an electric engraving were used to manually produce the artificial faults. While the natural damages were caused by using accelerated run-to-failure tests. The data collection process for both types of damages, i.e., artificial and real, was exposed under working conditions with different operating parameters such as loading torque, rotational speed and radial force. In total, the Paderborn datasets was collect under 6 different operating conditions including 3 conditions with artificial damages (denoted as domains I, J and K) and 3 conditions with real damages (denoted as domains L, M, and N). For example, the loading torque varies from 0.1 to 0.7 Nm and the radial force varies from 400 to 1000 N, while the rotational speed is fixed at 1500 RPM. Each operating condition (i.e., domain) contains three classes, namely, healthy class, inner fault (IF) class, and outer fault (OF) class. To prepare the data samples for the Paderborn dataset, we adopted sliding windows with a fixed length of 5,120 and a shifting size of 4,096 (Ragab et al. 2021). As such, we generated 12,340 for each artificial domain (i.e., I, J, and K) and 13,640 samples for each real domain (i.e., L, Mand N) respectively.

The process of automatic generalization is one of the elements of spatial data preparation for the purpose of creating digital cartographic studies. The presented data include a part of the process of generalization of building groups obtained from the Open Street Map databases (OSM) [1].

This repository contains images from the thesis and experimental result data, where the experimental result data is recorded in an Excel sheet with three columns containing the values of the model's training accuracy, testing accuracy, and generalizability assessment metrics, including IoU,IoU-B, Spectral Norm, EI, and Nuclear Norm.

A new dataset for natural language code search evaluating different types of generalization

Github: https://github.com/csi-greifflab/negative-class-optimization

Preprint: https://www.biorxiv.org/content/10.1101/2024.06.17.599333v1

A collection of 49 brain maps. Each brain map is a 3D array of values representing properties of the brain at different locations.

Collection description