Large Training Database Research Articles

Physics-informed artificial intelligence models for the seismic response prediction of rocking structures

Abstract The seismic response of a wide variety of structures, from small but irreplaceable museum exhibits to large bridge systems, is characterized by rocking. In addition, rocking motion is increasingly being used as a seismic protective strategy to limit the amount of seismic actions (moments) developed at the base of structures. However, rocking is a highly nonlinear phenomenon governed by non-smooth dynamic phases that make its prediction difficult. This study presents an alternative approach to rocking estimation based on a physics-informed convolutional neural network (PICNN). By training a group of PICNNs using limited datasets obtained from numerical simulations and encoding the known physics into the PICNNs, important predictive benefits are obtained relieving difficulties associated with over-fitting and minimizing the requirement for a large training database. Two models are created depending on the validation of the deep PICNN: the first model assumes that state variables including rotations and angular velocities are available, while the second model is useful when only acceleration measurements are known. The analysis is initiated by implementing K-means clustering. This is followed by a detailed statistical assessment and a comparative analysis of the response-histories of a rocking block. It is observed that the deep PICNN is capable of effectively estimating the seismic rocking response history when the rigid block does not overturn.

A mechanics-informed machine learning approach for modeling the elastoplastic behavior of fiber-reinforced composites

When machine learning (ML) techniques are used to predict the elastoplastic behavior of a fiber-reinforced composite, a large training database is typically required due to the complicated network architecture that is built to characterize the anisotropic plasticity. In this paper, a mechanics-informed ML approach that enables to employ a small training database is proposed to predict the elastoplastic behaviors of a unidirectional fiber-reinforced composite by incorporating mechanics-based decompositions of strain and stress into an artificial neural network (ANN). The built ANNs have simple structures and greatly enhance the prediction capability of the ML-based constitutive model when using a small database. The ML approach is further improved to predict the effect of the loading path. Direct numerical simulations (DNSs) based on a representative volume element are carried out to generate the datasets used for training and validating the ML-based constitutive model. By comparing the results obtained when using DNS and ML, it is shown that the proposed ML-based constitutive model offers excellent predictive accuracy even when using a small training database, and it provides much better results than a ML approach that does not include the decomposition of strain and stress.

Imaging inside highly scattering media using hybrid deep learning and analytical algorithm.

Imaging through highly scattering media is a challenging problem with numerous applications in biomedical and remote-sensing fields. Existing methods that use analytical or deep learning tools are limited by simplified forward models or a requirement for prior physical knowledge, resulting in blurry images or a need for large training databases. To address these limitations, we propose a hybrid scheme called Hybrid-DOT that combines analytically derived image estimates with a deep learning network. Our analysis demonstrates that Hybrid-DOT outperforms a state-of-the-art ToF-DOT algorithm by improving the PSNR ratio by 4.6 dB and reducing the resolution by a factor of 2.5. Furthermore, when compared to a deep learning stand-alone model, Hybrid-DOT achieves a 0.8 dB increase in PSNR, 1.5 times the resolution, and a significant reduction in the required dataset size (factor of 1.6-3). The proposed model remains effective at higher depths, providing similar improvements for up to 160 mean-free paths.

Machine learning techniques for chemical and type analysis of ocean oil samples via handheld spectrophotometer device

We designed and constructed a handheld, sturdy fluorescence spectrometry device for identifying samples from ocean oil spills. Two large training databases of autofluorescence spectra from raw oil samples (538 samples/1614 spectra and 767 samples/2301 spectra) were cross validated using support vector machine (SVM) to identify oil type and SARA (saturate, aromatic, resin, and asphaltene) contents. The device's performance was then validated on an independent set of 79 ocean oil samples, which were added to and then collected from ocean water during outdoor exposure to hot, humid weather to represent an actual oil spill. It successfully classified oil types with 92%–100% sensitivity and specificity and F1 scores of 85.7–100%. Further classification of light fuel oils into marine gas oil (MGO)-like and Bunker A (BA)-like categories was successful with the training set (raw oil samples), while less successful with the independent validation set (ocean oil samples). SARA content classification models performed well in training for the saturate (80.8% accuracy) and asphaltene (90.7%) contents. The developed training model was validated using ocean oil samples, and the resulting accuracies were 62.0% (saturate) and 93.7% (asphaltene). These results indicate the difficulties in classifying volatile light fuel oils with a low molecular weight that have experienced weathering effects, while high molecular weight compounds and general oil type can be analyzed.

Towards Multi-Domain Face Synthesis Via Domain-Invariant Representations and Multi-Level Feature Parts

Cross-domain face synthesis plays a positive role in the real world. It is challenging to synthesize high-quality faces across multiple domains based on limited paired data because the multiple mappings between different domains may interfere with each other. Cognitive science investigates that the brain can recognize the same person with multiple different expressions by extracting invariant information on the face and we humans perceive instances by decomposing them into parts. Motivated by these cognition, we propose a unified semi-supervised framework for multi-domain face synthesis by extracting a domain-invariant representation and exploiting parts of multi-level features. Specifically, realized by adversarial training with additional ability to utilize domain-specific information, a encoder is trained to remove domain-specific information and extract the domain-invariant representation from multiple inputs. Then, we utilize the multi-level feature parts extracted from inputs and reconstructed faces via a pre-trained recognition model to ensure that the domain-invariant representation contains enough useful semantic information. we also utilize the feature parts extracted from inputs and limited paired data to compose pseudo features in target domain for supervising the synthesis, which makes our framework suitable for large amounts of unpaired training data. By exploiting this framework, we can achieve face synthesis between multiple domains using some paired data together with a large training database without ground truth target faces. Experimental results demonstrate our framework achieves great performances on qualitative and quantitative evaluations under both artificial and uncontrolled environments, and our framework has competitive performances in single translation compared with specialized methods for translation between two specific domains.

A Deep Learning Approach for the Identification of Long-Duration Mixed Precipitation in Montréal (Canada)

ABSTRACT Long-duration mixed-precipitation events (freezing rain and/or ice pellets) are important cold-season hazards and understanding how climate change alters their occurrence is of high societal interest, particularly in urban areas. This study introduces a two-staged approach that employs deep learning to identify long-duration mixed precipitation over the Montréal area (Quebec, Canada) in archived climate model data using large-scale pressure patterns. The dominant dynamic mechanism leading to mixed precipitation in Montréal is pressure-driven channelling of winds along the St. Lawrence River Valley. A convolutional neural network (CNN) identifies the corresponding synoptic pattern by using a large training database derived from an ensemble of the Canadian Regional Climate Model, version 5 (CRCM5). The CRCM5 uses the diagnostic method of Bourgouin (2000) to simulate mixed precipitation and delivers training examples and corresponding class affiliations (labels) for this supervised classification task. The CNN correctly identifies more than 80% of the Bourgouin mixed-precipitation cases. In the next stage, the CNN is combined with temperature and precipitation conditions, which consider important preconditions for mixed precipitation and improve the performance of the approach. The evaluation of a CRCM5 run driven by ERA-Interim reanalysis data gives a Matthews correlation coefficient of 0.50. The deep learning approach can be applied to ensembles of regional climate models on the North American grid of the Coordinated Regional Downscaling Experiment (CORDEX-NA).

An Overview of AI-Assisted Design-on-Simulation Technology for Reliability Life Prediction of Advanced Packaging.

Several design parameters affect the reliability of wafer-level type advanced packaging, such as upper and lower pad sizes, solder volume, buffer layer thickness, and chip thickness, etc. Conventionally, the accelerated thermal cycling test (ATCT) is used to evaluate the reliability life of electronic packaging; however, optimizing the design parameters through ATCT is time-consuming and expensive, reducing the number of experiments becomes a critical issue. In recent years, many researchers have adopted the finite-element-based design-on-simulation (DoS) technology for the reliability assessment of electronic packaging. DoS technology can effectively shorten the design cycle, reduce costs, and effectively optimize the packaging structure. However, the simulation analysis results are highly dependent on the individual researcher and are usually inconsistent between them. Artificial intelligence (AI) can help researchers avoid the shortcomings of the human factor. This study demonstrates AI-assisted DoS technology by combining artificial intelligence and simulation technologies to predict wafer level package (WLP) reliability. In order to ensure reliability prediction accuracy, the simulation procedure was validated by several experiments prior to creating a large AI training database. This research studies several machine learning models, including artificial neural network (ANN), recurrent neural network (RNN), support vector regression (SVR), kernel ridge regression (KRR), K-nearest neighbor (KNN), and random forest (RF). These models are evaluated in this study based on prediction accuracy and CPU time consumption.

The influence of training data variability on a supervised machine learning classifier for Structure from Motion (SfM) point clouds of rock slopes

Supervised Machine Learning (ML) can be used to automatically interpret remote sensing data in engineering geology, with applications for rockfall and landslide characterization. However, supervised algorithms typically require very large training databases from which to learn predictive relationships, and there is little guidance on how to construct such a database in an earth science context with high temporal and spatial heterogeneity. This study builds a supervised classifier to perform basic rock slope characterization on Structure from Motion (SfM) point clouds collected under a variety of conditions. Eight datasets were collected in Colorado and Utah, USA, with multiple different sensor platforms (terrestrial and aerial), rock types, seasons, and lighting conditions, and each dataset was manually labeled to identify regions of vegetation, rock, soil, talus, and snow. A total of 2560 Random Forest models were built with different combinations of training datasets and combinations of geometric and color features. Most models were able to identify vegetation and rock with high accuracy (median F scores of 86% and 68% respectively), but performance for soil, talus, and snow was overall much poorer, and the median overall accuracy of generalized classifiers was 60%. Many characteristics of the training data were found to have significant effects on generalization accuracy, indicating that training datasets must be curated to be applicable to specific data collection parameters, seasons, lighting conditions, and geological settings. We conclude that high accuracy generalized results can be obtained, but the ML model must be carefully constructed, and its limitations acknowledged.

Facial masks and soft‐biometrics: Leveraging face recognition CNNs for age and gender prediction on mobile ocular images

We address the use of selfie ocular images captured with smartphones to estimate age and gender. Partial face occlusion has become an issue due to the mandatory use of face masks. Also, the use of mobile devices has exploded, with the pandemic further accelerating the migration to digital services. However, state-of-the-art solutions in related tasks such as identity or expression recognition employ large Convolutional Neural Networks, whose use in mobile devices is infeasible due to hardware limitations and size restrictions of downloadable applications. To counteract this, we adapt two existing lightweight CNNs proposed in the context of the ImageNet Challenge, and two additional architectures proposed for mobile face recognition. Since datasets for soft-biometrics prediction using selfie images are limited, we counteract over-fitting by using networks pre-trained on ImageNet. Furthermore, some networks are further pre-trained for face recognition, for which very large training databases are available. Since both tasks employ similar input data, we hypothesise that such strategy can be beneficial for soft-biometrics estimation. A comprehensive study of the effects of different pre-training over the employed architectures is carried out, showing that, in most cases, a better accuracy is obtained after the networks have been fine-tuned for face recognition.

Fast Strain Estimation and Frame Selection in Ultrasound Elastography Using Machine Learning.

Ultrasound elastography aims to determine the mechanical properties of the tissue by monitoring tissue deformation due to internal or external forces. Tissue deformations are estimated from ultrasound radio frequency (RF) signals and are often referred to as time delay estimation (TDE). Given two RF frames I1 and I2 , we can compute a displacement image, which shows the change in the position of each sample in I1 to a new position in I2 . Two important challenges in TDE include high computational complexity and the difficulty in choosing suitable RF frames. Selecting suitable frames is of high importance because many pairs of RF frames either do not have acceptable deformation for extracting informative strain images or are decorrelated and deformation cannot be reliably estimated. Herein, we introduce a method that learns 12 displacement modes in quasi-static elastography by performing principal component analysis (PCA) on displacement fields of a large training database. In the inference stage, we use dynamic programming (DP) to compute an initial displacement estimate of around 1% of the samples and then decompose this sparse displacement into a linear combination of the 12 displacement modes. Our method assumes that the displacement of the whole image could also be described by this linear combination of principal components. We then use the GLobal Ultrasound Elastography (GLUE) method to fine-tune the result yielding the exact displacement image. Our method, which we call PCA-GLUE, is more than 10× faster than DP in calculating the initial displacement map while giving the same result. This is due to converting the problem of estimating millions of variables in DP into a much simpler problem of only 12 unknown weights of the principal components. Our second contribution in this article is determining the suitability of the frame pair I1 and I2 for strain estimation, which we achieve by using the weight vector that we calculated for PCA-GLUE as an input to a multilayer perceptron (MLP) classifier. We validate PCA-GLUE using simulation, phantom, and in vivo data. Our classifier takes only 1.5 ms during the testing phase and has an F1-measure of more than 92% when tested on 1430 instances collected from both phantom and in vivo data sets.

NEUTROSOPHIC TEXTURE-REGION DIFFERENCE-BASED FUZZY c-MEANS CLUSTERING OF ULTRASOUND TUMOR IMAGES

Owing to its cost effective and non-ionizing procedure, ultrasound imaging has become a primary tool for detection and diagnosis of breast tumors compared to mammography or biopsy. But the precise interpretation of the images obtained is limited due to the intensity inhomogeneties, speckle noise, low contrast, etc. present in the observed images. Also, tumors can appear in different echogenic patterns including visually challenging to detect isoechoic and slightly hypoechoic tumors, and distinguishing the tumor by means of intensity variations only will not suffice. Therefore it is important to preprocess the image to remove unwanted noise and variations, and thereby provide adequate enhancement to highlight the region of interest. In this paper, a segmentation method is proposed where the query images are clustered using fuzzy c-means clustering (FCM) after neutrosophic filtering and enhancement followed by extraction of textural and edge features through the use of Gabor and compass filters. In this work, Prewitt compass filters were implemented to capture the maximum region difference between the tumor and background. By using filters of different orientations, sufficient texture and edge information were able to be captured by using the combination of Gabor and compass filters, which can then aid in the accurate detection of isoechoic and hypoechoic tumors. The neutrosophic preprocessing implemented in this work models the ultrasound noise as combination of Gaussian and Rayleigh distributions, and shows better signal-to-noise ratio than traditionally modeling speckle as Gaussian alone. The proposed method gave an increase of 42.25%, 40.15%, 81.32% and 50.15% and a decrease of 95.83% compared to standard Otsu thresholding, an increase of 23.34%, 19.57%, 24.53% and 24.71% and a decrease of 91.81% compared to simple FCM and an increase of 37.89%, 44.64%, 35.86% and 48.11% and a decrease of 93.68% compared to active contour segmentation in structural similarity index, Jaccard index, Dice coefficient and accuracy and mean square error, respectively. On contrasting with recent learning-based methodologies for tumor segmentation, the proposed method gave comparable accuracy without the need for large training database and expensive graphical processing unit (GPU) computing capabilities.

0442 Experimenting Automatic Sleep Analysis Application in a Clinical Context

Abstract Introduction Multiplication of publications describing groundbreaking automatic sleep analysis processes and algorithms push for real-life experimentation in clinical context, outside of controlled research environments. Methods Various automatic sleep analysis processes from the literature were implemented and orchestrated in a streamlined workflow. Artificial Intelligence algorithms using regular statistical learning or deep learning were re-trained on our own data after repeating the ad-hoc pre-processing steps described in the corresponding articles. For this, we used polysomnographic records previously taped in our clinic, subject to adequate legal authorizations and agreements: 500 nights from single patients with various pathologies. Those trained models were then applied to newly recorded polysomnographies through a platform developed and hosted on premise. For each polysomnography, a standardized and automatized report were generated and transmitted to the clinician in charge of the analysis. This report contains algorithms outputs, including automatic staging and related statistics such as hypnodensity, quantitative electroencephalography (EEG) analysis, spindles detection and automatic diagnosis. Aggregated record statistics are displayed next to our database statistics for benchmarking purposes. Results For sleep staging, we not only reproduced the results of the selected literature but obtained better metrics: a 0.76 Kappa agreement vs 0.69 in the literature. This may be due to our larger training database or the quality of physiologic signals in our data. Clinicians showed interest in the automatic staging part of the analysis. They noticed algorithm errors are mostly focused on ambiguous epochs, just like visual scoring. However, they found help into automated output and explanatory variables (hypnodensity) to score those ambiguous epochs. Conclusion Automatic sleep analysis algorithms used as decision helping tools shows real potential and should be generalized, as long as underlying processes are published and understood by users and clinicians. Support Banque Publique d’Investissement.

From Optimization-Based Machine Learning to Interpretable Security Rules for Operation

Various supervised machine learning approaches have been used in the past to assess the power system security (also known as reliability). This is typically done by training a classifier on a large number of operating points whose postfault status (stable or unstable) has been determined via time-domain simulations. The output of this training process can be expressed as a security rule that is used online to classify an operating point. A critical, and little-studied aspect of these approaches is the interpretability of the rules produced. The lack of interpretability is a well-known issue of some machine learning approaches, especially when dealing with difficult classification problems. In the case of the security assessment of the power system, which is a complex mission-critical task, interpretability is a key requirement for the adoption and deployment by operators of these approaches. In this paper, for the first time, we explore the tradeoff between predictive accuracy and interpretability in the context of power system security assessment. We begin by demonstrating how decision trees (DTs) can be used to learn data-driven security rules and use the tree depth as a measure for interpretability. We leverage disjunctive programming to formulate novel training methods, capable of learning high-quality DTs while still maintaining interpretability. In particular, we propose two new approaches: 1) optimal classification trees is proposed for training DTs of low-depth and 2) greedy optimization-based tree is proposed for training DTs of intermediate depth, where the increased computational burden is managed by exploiting the nested tree structure. We also demonstrate that the ability to generate high-quality interpretable rules can actually translate to impressive benefits in terms of training requirements. Through case studies on the IEEE 68-bus system, we demonstrate that the proposed methods can produce DTs of higher quality compared to the state-of-the-art approach classification and regression tree approach, also if the DT was trained on a significant smaller database, resulting in computational savings of 80%. Given that generating a large training database is a practical bottleneck in these data-driven approaches, this is a significant breakthrough for real-world application.

Application of the Least Square Support Vector Machine for point-to-point forecasting of the PV Power

<p>In today's industrial world, the growing capacity of renewable energy sources is a crucial factor for sustainable power generation. The application of solar photovoltaic (PV) energy sources, as a clean and safe renewable energy resource has found great attention among the consumers in the recent decades. Accurate forecasting of the generated PV power is an important task for scheduling the generators and planning the consumption patterns of customers to save electricity costs. To this end, it is necessary to develop a global model of the generated power based on the effective factors which are mainly the solar radiation intensity and the ambient weather temperature. As a result of the wide numerical range of these parameters and various weather conditions, a large training database must be used for developing the models, which results in high-computational complexity of the algorithms used for training the models. In this paper, a novel algorithm for point to point prediction of the generated power based on the least squares support vector machine (LS-SVM) has been proposed which can handle the large training database with a very fewer deal of computation and benefits from reasonable accuracy and generalization capability. </p>

Abstract MS1-2: Deep learning systems for reading mammograms and breast tomosynthesis

Abstract Recent breakthroughs in machine learning resulted in powerful techniques that allow reseachers to accelerate development of automated systems for interpreting medical images. With the use of deep learning and very large training databases, limitations of conventional computer aided detection (CAD) systems, which are widely used in mammography, can be overcome. It is expected that in the coming years algorithms will become more accurate that human readers for interpreting screening mammograms. This will enable physicians to substantially increase quality of mammography. It also will reduce the need for human resources, which is important because many countries face challenges with recruitment of mammographers. In the presentation, experimental study results will be presented which illustrate that deep learning systems are approaching the level of performance of human readers and that there is potential to increase the quality of automated interpretation of mammograms and breast tomosynthesis beyond to the level of experts Citation Format: Karssemeijer N. Deep learning systems for reading mammograms and breast tomosynthesis [abstract]. In: Proceedings of the 2018 San Antonio Breast Cancer Symposium; 2018 Dec 4-8; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2019;79(4 Suppl):Abstract nr MS1-2.

Subject-Specific Convolutional Neural Networks for Accelerated Magnetic Resonance Imaging.

Magnetic Resonance Imaging (MRI) is one of the leading modalities for medical imaging, providing excellent soft-tissue contrast without exposure to ionizing radiation. Despite continuing advances in MRI, long scan times remain a major limitation in clinical applications. Parallel imaging is a technique for scan time acceleration in MRI, which utilizes the spatial variations in the reception profiles of receiver coil arrays to reconstruct images from undersampled Fourier space, i.e. k-space. One of the most commonly used parallel imaging techniques employs interpolation of missing k-space information by using linear shift-invariant convolutional kernels. These kernels are trained on a limited amount of autocalibration signal (ACS) for each scan. We propose a novel method for parallel imaging, Robust Artificial-neural-networks for k-space Interpolation (RAKI), which uses scan-specific convolutional neural networks (CNNs) to perform improved k-space interpolation. Three-layer CNNs are trained using only scan-specific ACS data, alleviating the need for large training databases. The proposed method was tested in ultra-high resolution brain MRI and quantitative cardiac MRI, acquired with various acceleration rates. Improved noise resilience as compared to existing parallel imaging methods was observed for high acceleration rates or in the presence of low signal-to-noise ratio (SNR). Furthermore, RAKI successfully reconstructed images for quantitative cardiac MRI, even when using the same CNN across images with varying contrasts. These results indicate that RAKI achieves improved noise performance without overfitting to specific image contents, and offers great promise for improved acceleration in a wide range of MRI applications.

Active graph based semi-supervised learning using image matching: Application to handwritten digit recognition

With the availability of large amounts of documents and multimedia content to be classified, the creation of new databases with labeled examples is an expensive task. Efficient supervised classifiers often require large training databases that are not always immediately available. Active learning approaches solve this issue by querying an expert to set a label to particular instances. In this paper, we present a novel active learning strategy for the classification of handwritten digits. The proposed method is based on a k-nearest neighbor graph obtained with an image deformation model, which takes into account local deformations. During the active learning procedure, the user is first asked to label the vertices with the highest number of neighbors. Thus, the expert sets the label to the examples that are more likely to propagate their labels to a high number of close neighbors. Then, a label propagation function is performed to automatically label the examples. The procedure is repeated until all the images are labeled. We evaluate the performance of the method on four databases corresponding to different scripts (Latin, Bangla, Devnagari, and Oriya). We show that it is possible to label only 332 images in the MNIST training database to obtain an accuracy of 98.54% on this same database (60000 images). The robustness of the method is highlighted by the performance of handwritten digit recognition in different scripts.

Improving Retrieval Efficacy of Homology Searches Using the False Discovery Rate.

Over the past few decades, discovery based on sequence homology has become a widely accepted practice. Consequently, comparative accuracy of retrieval algorithms (e.g., BLAST) has been rigorously studied for improvement. Unlike most components of retrieval algorithms, the E-value threshold criterion has yet to be thoroughly investigated. An investigation of the threshold is important as it exclusively dictates which sequences are declared relevant and irrelevant. In this paper, we introduce the false discovery rate (FDR) statistic as a replacement for the uniform threshold criterion in order to improve efficacy in retrieval systems. Using NCBI's BLAST and PSI-BLAST software packages, we demonstrate the applicability of such a replacement in both non-iterative (BLASTFDR) and iterative (PSI-BLAST(FDR)) homology searches. For each application, we performed an evaluation of retrieval efficacy with five different multiple testing methods on a large training database. For each algorithm, we choose the best performing method, Benjamini-Hochberg, as the default statistic. As measured by the threshold average precision, BLAST(FDR) yielded 14.1 percent better retrieval performance than BLAST on a large (5,161 queries) test database and PSI-BLAST(FDR) attained 11.8 percent better retrieval performance than PSI-BLAST. The C++ source code specific to BLAST(FDR) and PSI-BLAST(FDR) and instructions are available at http://www.cs.mtsu.edu/~hcarroll/blast_fdr/.

A novel multiple-instance learning-based approach to computer-aided detection of tuberculosis on chest X-rays.

To reach performance levels comparable to human experts, computer-aided detection (CAD) systems are typically optimized following a supervised learning approach that relies on large training databases comprising manually annotated lesions. However, manually outlining those lesions constitutes a difficult and time-consuming process that renders detailedly annotated data difficult to obtain. In this paper, we investigate an alternative approach, namely multiple-instance learning (MIL), that does not require detailed information for optimization. We have applied MIL to a CAD system for tuberculosis detection. Only the case condition (normal or abnormal) was required during training. Based upon the well-known miSVM technique, we propose an improved algorithm that overcomes miSVM's drawbacks related to positive instance underestimation and costly iteration. To show the advantages of our MIL-based approach as compared with a traditional supervised one, experiments with three X-ray databases were conducted. The area under the receiver operating characteristic curve was utilized as a performance measure. With the first database, for which training lesion annotations were available, our MIL-based method was comparable to the supervised system ( 0.86 versus 0.88 ). When evaluating the remaining databases, given their large difference with the previous image set, the most appealing strategy was to retrain the CAD systems. However, since only the case condition was available, only the MIL-based system could be retrained. This scenario, which is common in real-world applications, demonstrates the better adaptation capabilities of the proposed approach. After retraining, our MIL-based system significantly outperformed the supervised one ( 0.86 versus 0.79 and 0.91 versus 0.85 , and p=0.0002 , respectively).

A reduced probabilistic neural network for the classification of large databases

The probabilistic neural network (PNN) is a special type of radial basis neural network used mainly for classification problems. Due to the size of the network after training, this type of network is usually used for problems with a small-sized training dataset. In this paper, a new training algorithm is presented for use with large training databases. Application to the handwritten digit database shows that the reduced PNN performs better than the standard PNN for all of the studied cases with a big gain in size and processing speed. This new type of neural network can be used easily for problems with large training databases like biometrics and data mining applications. An extension of the network is possible for new training samples and/or classes without retraining.