Need For Large Training Sets Research Articles

Transferability of Machine Learning Models for Geogenic Contaminated Groundwaters.

Machine learning models show promise in identifying geogenic contaminated groundwaters. Modeling in regions with no or limited samples is challenging due to the need for large training sets. One potential solution is transferring existing models to such regions. This study explores the transferability of high fluoride groundwater models between basins in the Shanxi Rift System, considering six factors, including modeling methods, predictor types, data size, sample/predictor ratio (SPR), predictor range, and data informing. Results show that transferability is achieved only when model predictors are based on hydrochemical parameters rather than surface parameters. Data informing, i.e., adding samples from challenging regions to the training set, further enhances the transferability. Stepwise regression shows that hydrochemical predictors and data informing significantly improve transferability, while data size, SPR, and predictor range have no significant effects. Additionally, despite their stronger nonlinear capabilities, random forests and artificial neural networks do not necessarily surpass logistic regression in transferability. Lastly, we utilize the t-SNE algorithm to generate low-dimensional representations of data from different basins and compare these representations to elucidate the critical role of predictor types in transferability.

Segmenting computed tomograms for cardiac ablation using machine learning leveraged by domain knowledge encoding.

Segmentation of computed tomography (CT) is important for many clinical procedures including personalized cardiac ablation for the management of cardiac arrhythmias. While segmentation can be automated by machine learning (ML), it is limited by the need for large, labeled training data that may be difficult to obtain. We set out to combine ML of cardiac CT with domain knowledge, which reduces the need for large training datasets by encoding cardiac geometry, which we then tested in independent datasets and in a prospective study of atrial fibrillation (AF) ablation. We mathematically represented atrial anatomy with simple geometric shapes and derived a model to parse cardiac structures in a small set of N = 6 digital hearts. The model, termed "virtual dissection," was used to train ML to segment cardiac CT in N = 20 patients, then tested in independent datasets and in a prospective study. In independent test cohorts (N = 160) from 2 Institutions with different CT scanners, atrial structures were accurately segmented with Dice scores of 96.7% in internal (IQR: 95.3%-97.7%) and 93.5% in external (IQR: 91.9%-94.7%) test data, with good agreement with experts (r = 0.99; p < 0.0001). In a prospective study of 42 patients at ablation, this approach reduced segmentation time by 85% (2.3 ± 0.8 vs. 15.0 ± 6.9 min, p < 0.0001), yet provided similar Dice scores to experts (93.9% (IQR: 93.0%-94.6%) vs. 94.4% (IQR: 92.8%-95.7%), p = NS). Encoding cardiac geometry using mathematical models greatly accelerated training of ML to segment CT, reducing the need for large training sets while retaining accuracy in independent test data. Combining ML with domain knowledge may have broad applications.

The Devil is in the Upsampling: Architectural Decisions Made Simpler for Denoising with Deep Image Prior.

Deep Image Prior (DIP) shows that some network architectures inherently tend towards generating smooth images while resisting noise, a phenomenon known as spectral bias. Image denoising is a natural application of this property. Although denoising with DIP mitigates the need for large training sets, two often intertwined practical challenges need to be overcome: architectural design and noise fitting. Existing methods either handcraft or search for suitable architectures from a vast design space, due to the limited understanding of how architectural choices affect the denoising outcome. In this study, we demonstrate from a frequency perspective that unlearnt upsampling is the main driving force behind the denoising phenomenon with DIP. This finding leads to straightforward strategies for identifying a suitable architecture for every image without laborious search. Extensive experiments show that the estimated architectures achieve superior denoising results than existing methods with up to 95% fewer parameters. Thanks to this under-parameterization, the resulting architectures are less prone to noise-fitting.

UWB Radar Applied to Lane Occupation and Vehicle Classification

This article describes the use of a commercial UWB radar for vehicle classification and lane occupation detection using real-world data acquired in an urban environment. We compare two radar image processing schemes: one based on deep learning using raw data produced by the radar, and a second method employing traditional machine learning algorithms using features extracted from raw data. We verify experimentally that both schemes lead to reasonably accurate estimates without the need of large training sets.

Impact of Training Set Size on the Ability of Deep Neural Networks to Deal with Omission Noise

Deep Learning usually requires large amounts of labeled training data. In remote sensing, deep learning is often applied for land cover and land use classification as well as street network and building segmentation. In case of the latter, a common way of obtaining training labels is to leverage crowdsourced datasets which can provide numerous types of spatial information on a global scale. However, labels from crowdsourced datasets are often limited in the sense that they potentially contain high levels of noise. Understanding how such noisy labels impede the predictive performance of Deep Neural Networks (DNNs) is crucial for evaluating if crowdsourced data can be an answer to the need for large training sets by DNNs. One way towards this understanding is to identify the factors which affect the relationship between label noise and predictive performance of a model. The size of the training set could be one of these factors since it is well known for being able to greatly influence a model’s predictive performance. In this work we pick the size of the training set and study its influence on the robustness of a model against a common type of label noise known as omission noise. To this end, we utilize a dataset of aerial images for building segmentation and create several versions of the training labels by introducing different amounts of omission noise. We then train a state-of-the-art model on subsets of varying size of those versions. Our results show that the training set size does play a role in affecting the robustness of our model against label noise: A large training set improves the robustness of our model against omission noise.

Predicting the dissolution kinetics of silicate glasses by topology-informed machine learning

Machine learning (ML) regression methods are promising tools to develop models predicting the properties of materials by learning from existing databases. However, although ML models are usually good at interpolating data, they often do not offer reliable extrapolations and can violate the laws of physics. Here, to address the limitations of traditional ML, we introduce a “topology-informed ML” paradigm—wherein some features of the network topology (rather than traditional descriptors) are used as fingerprint for ML models—and apply this method to predict the forward (stage I) dissolution rate of a series of silicate glasses. We demonstrate that relying on a topological description of the atomic network (i) increases the accuracy of the predictions, (ii) enhances the simplicity and interpretability of the predictive models, (iii) reduces the need for large training sets, and (iv) improves the ability of the models to extrapolate predictions far from their training sets. As such, topology-informed ML can overcome the limitations facing traditional ML (e.g., accuracy vs. simplicity tradeoff) and offers a promising route to predict the properties of materials in a robust fashion.

Brain haemorrhage detection using a SVM classifier with electrical impedance tomography measurement frames.

Brain haemorrhages often require urgent treatment with a consequent need for quick and accurate diagnosis. Therefore, in this study, we investigate Support Vector Machine (SVM) classifiers for detecting brain haemorrhages using Electrical Impedance Tomography (EIT) measurement frames. A 2-layer model of the head, along with a series of haemorrhages, is designed as both numerical models and physical phantoms. EIT measurement frames, taken from an electrode array placed on the head surface, are used to train and test linear SVM classifiers. Various scenarios are implemented on both platforms to examine the impact of variables such as noise level, lesion location, lesion size, variation in electrode positioning, and variation in anatomy, on the classifier performance. The classifier performed well in numerical models (sensitivity and specificity of 90%+) with signal-to-noise ratios of 60 dB+, was independent of lesion location, and could detect lesions reliably down to the tested minimum volume of 5 ml. Slight variations in electrode layout did not affect performance. Performance was affected by variations in anatomy however, emphasising the need for large training sets covering different anatomies. The phantom models proved more challenging, with maximal sensitivity and specificity of 75% when used with the linear SVM. Finally, the performance of two more complex classifiers is briefly examined and compared to the linear SVM classifier. These results demonstrate that a radial basis function (RBF) SVM classifier and a neural network classifier can improve detection accuracy. Classifiers applied to EIT measurement frames is a novel approach for lesion detection and may offer an effective diagnostic tool clinically. A challenge is to translate the strong results from numerical models into real world phantoms and ultimately human patients, as well as the selection and development of optimal classifiers for this application.