Unlearning from experience to avoid spurious correlations

In the development of social economy and scientific and technological innovation, the image processing mode and classification model chosen by network technology platform is becoming more and more changeable, but in essence, it is necessary to obtain characteristic information in effective image recognition and choose high-quality network algorithm and processing technology to complete image processing and image classification. Therefore, on the basis of understanding the current research trend of computer image processing and image classification model methods, this paper conducts in-depth discussion on the image processing methods and image classification model training design with artificial intelligence as the core and takes the image classification model of transfer learning as an example for practical exploration. The final results show that the image processing method and image classification model based on artificial intelligence have strong performance advantages in practical application.KeywordsArtificial intelligenceImage processingImage classificationThe migration study

Deep-learning (DL) methods have recently been introduced for seismic signal processing. Using DL methods, many researchers have adopted these novel techniques in an attempt to construct a DL model for seismic data reconstruction. The performance of DL-based methods depends heavily on what is learned from the training data. We focus on constructing the DL model that well reflect the features of target data sets. The main goal is to integrate DL with an intuitive data analysis approach that compares similar patterns prior to the DL training stage. We have developed a two-sequential method consisting of two stages: (1) analyzing training and target data sets simultaneously for determining the target-informed training set and (2) training the DL model with this training data set to effectively interpolate the seismic data. Here, we introduce the convolutional autoencoder t-distributed stochastic neighbor embedding (CAE t-SNE) analysis that can provide the insight into the results of interpolation through the analysis of training and target data sets prior to DL model training. Our method was tested with synthetic and field data. Dense seismic gathers (e.g., common-shot gathers) were used as a labeled training data set, and relatively sparse seismic gathers (e.g., common-receiver gathers [CRGs]) were reconstructed in both cases. The reconstructed results and signal-to-noise ratios demonstrated that the training data can be efficiently selected using CAE t-SNE analysis, and the spatial aliasing of CRGs was successfully alleviated by the trained DL model with this training data, which contain target features. These results imply that the data analysis for selecting target-informed training set is very important for successful DL interpolation. In addition, our analysis method can also be applied to investigate the similarities between training and target data sets for other DL-based seismic data reconstruction tasks.

This dataset is composed of annotations of the five hemorrhage subtypes (subarachnoid, intraventricular, subdural, epidural, and intraparenchymal hemorrhage) typically encountered at brain CT.

The detection of mine-like objects (MLOs) in sidescan sonar (SSS) imagery continues to be a challenging task. In practice, subject matter experts tediously analyze images searching for MLOs. In the literature, there are many attempts at automated target recognition (ATR) to detect the MLOs. This paper focuses on the classifiers that use computer vision and machine learning approaches. These techniques require large amounts of data, which is often prohibitive. For this reason, the use of synthetic and semisynthetic data sets for training and testing is commonplace. This paper shows how a simple semisynthetic data creation scheme can be used to pretest these data-hungry training algorithms to determine what features are of value. The paper provides real-world testing and training data sets in addition to the semisynthetic training and testing data sets. The paper considers the Haar-like and local binary pattern (LBP) features with boosting, showing improvements in performance with real classifiers over semisynthetic classifiers and improvements in performance as semisynthetic data set size increases.

Fused deep neural networks for sustainable and computational management of heat-transfer pipeline diagnosis

A probabilistic topic model using deep visual word representation for simultaneous image classification and annotation

The predictions of text classifiers are often driven by spurious correlations – e.g., the term “Spielberg” correlates with positively reviewed movies, even though the term itself does not semantically convey a positive sentiment. In this paper, we propose a method to distinguish spurious and genuine correlations in text classification. We treat this as a supervised classification problem, using features derived from treatment effect estimators to distinguish spurious correlations from “genuine” ones. Due to the generic nature of these features and their small dimensionality, we find that the approach works well even with limited training examples, and that it is possible to transport the word classifier to new domains. Experiments on four datasets (sentiment classification and toxicity detection) suggest that using this approach to inform feature selection also leads to more robust classification, as measured by improved worst-case accuracy on the samples affected by spurious correlations.

Industrial Revolution (IR) improves the way we live, work and interact with each other by using state of the art technologies. IR-4.0 describes a future state of industry which is characterized through the digitization of economic and production flows. The nine pillars of IR-4.0 are dependent on Big Data Analytics, Artificial Intelligence, Cloud Computing Technologies and Internet of Things (IoT). Image datasets are most valuable among other types of Big Data. Image Classification Models (ICM) are considered as an appropriate solution for Business Intelligence. However, due to complex image characteristics, one of the most critical issues encountered by the ICM is the Concept Drift (CD). Due to CD, ICM are not able to adapt and result in performance degradation in terms of accuracy. Therefore, ICM need better adaptability to avoid performance degradation during CD. Adaptive Convolutional ELM (ACNNELM) is one of the best existing ICM for handling multiple types of CD. However, ACNNELM does not have sufficient adaptability. This paper proposes a more autonomous adaptability module, based on Meta-Cognitive principles, for ACNNELM to further improve its performance accuracy during CD. The Meta-Cognitive module will dynamically select different CD handling strategies, activation functions, number of neurons and restructure ACNNELM as per changes in the data.

Most popular image classification methods mainly focus on classification ability rather than recognizing new things. However, human lay emphasis on cognition first and then classification, which is closely related to human memory system. Though many memory models have been proposed, they are studied in word list whereas the reports about natural images are still limited. This paper proposes a memory model for image recognition and classification based on convolutional neural network and Bayesian decision. First the image feature is extracted by convolutional neural network and stored in binary form. Then the representation, storage and retrieval processes of visual images are modeled. The test image feature vector is matched in parallel to the studied image vectors, and the likelihood values are calculated. Finally, the odd that the test image belongs to a new class is computed based on all likelihood values. If the odd value is greater than a certain threshold, the test image is regarded as new; otherwise, the Bayesian decision rule for image classification is performed. Experimental results on Caltech-101 and Caltech-256 datasets show that the proposed method can perform well in image recognition and classification tasks. And the hit probability of the method is higher than two typical methods, SRC and ELM, at present while the false alarm rate is far lower than them.

To solve the main problem of recognition accuracy, many image classification models have been implemented. A lot of attention was paid to Machine Learning. In this work, we will examine the problem of image classification related on transmission training to study whether it will work better in point of accuracy and efficiency with new sets of image data through Transfer Learning. Transfer Learning is a method of using the knowledge of a pre-trained model in another task. In this article, we will compare the image classification results of Logistic Regression (LR), Linear SVM and Random Forest Classifiers (RFC) using the pre – trained VGG-16 model. Image classification problem is implemented using Caltech - 101 and Flowers - 17 datasets.KeywordsTransfer learningFeature extractionVGG-16 modelLogistic regressionSVM and random forest classifiersImage classification

A renewed and growing interest in phenotypic drug screening approaches in the field of drug discovery is observed, as it has become apparent that target-oriented drug discovery assays have inherent limitations and cannot fulfil the urgent unmet medical need for novel drugs. The shortcomings of target-oriented drug screening assays are especially apparent in the field of antibiotic drug discovery, where target-based approaches largely failed to translate screening hits to clinically relevant drugs. 
\nIn this thesis, a proteomics-based phenotypic drug screening approach using MALDI-TOF mass spectrometry was developed, which is able to detect sub-lethal stress in bacterial cells provoked by antibiotics. To achieve this, mass spectra of whole-cells exposed to known antibiotics at concentrations below the minimal inhibitory concentration (MIC) were used to extract relevant mass spectral peaks with a data-dependent and automated computational pipeline created in the MATLAB environment. Using the selected subset of mass spectral peaks, classification models were trained to recognize general mass spectral responses provoked by unknown drugs in the cellular proteome. Additionally, the classification models proved capable of identifying the mechanisms of action of unknown drugs.
\nTo establish and validate the best performing classification modeling procedure, four different feature selection algorithms and nine classification models were analyzed in detail using an Escherichia coli data set composed of over 900 spectra, involving 17 antibiotics with four different mechanisms of action, at concentrations ranging 1×MIC down to 1/32×MIC in a two-fold dilution series. Four different feature selection approaches were investigated to ensure the extraction of relevant mass spectral data in response to the different antibiotics for classification modeling. The selection approaches included (1) a random forest of decision trees, (2) sequential forward feature selection, and (3) sequential backward feature selection. Mass spectral peaks selected by two or all three of these feature selection approaches were combined into (4) an aggregated feature set. Classification models were trained for all combinations of nine model types and the four feature sets. In this thesis two classification problems were investigated. First, a binary classification problem, to differentiate between affected cells, and non-affected cells based on selected mass spectral peaks. Second, a multi-class model was trained to detect and distinguish between the different antibiotic mechanisms of action, a highly desired drug screening assay characteristic. The combination of these elements yielded 72 models, which were evaluated based on their overall classification accuracy. The overall classification accuracy was determined using internal 10-fold cross-validation and external validation, which was performed with a blind set of 20 drugs. The internal and external validation studies showed that the aggregated feature set in combination with a quadratic support vector machine-based model (Q-SVM) resulted in the best classification performance. For the E. coli data set, this was represented by an overall accuracy of 0.92 for internal validation and an accuracy of 0.95 for the external validation of the Q-SVM model. Classifying based on the mechanism of action of the antibiotics resulted in a classification accuracy of 0.67 for internal validation and 0.80 for external validation. Furthermore, it was shown that the peak selection method was able to identify relevant, known stress associated proteins within the aggregated feature sets of both the binary and the mechanism of action model. 
\nAfter the experimental workflow and the computational pipeline were established based on E. coli data, the method was applied to four different organisms (the Gram-positive bacterium Staphylococcus aureus, the fungi Saccharomyces cerevisiae and Candida albicans, and human HeLa cancer cell line) and different proteomic responses, to explore the versatility and transferability of the developed screening assay. The applicability of the method was demonstrated by the consistent performance of the classification models generated with the experimental and computational pipeline. This resulted in binary model accuracies between 0.92 and 0.97 for internal and 0.77 and 0.95 for external validation, depending on the assayed organism and data set complexity. For mechanism of action models, model accuracies ranged between 0.73 and 0.96 for internal and 0.66 and 0.93 for external validation.
\nThe application of the developed assay on different organisms with different drug stressors highlighted several advantageous characteristics of the developed MALDI-TOF MS screening approach. Both the binary and mechanism of action classification models of S. aureus correctly identified an antibiotic drug (fusidic acid) in the blind test set, which had a target binding activity that was not present in the training data set. This implicates the ability of the method to detect novel drugs within known global mechanism of action for which the model was trained. Moreover, external validation of S. cerevisiae showed that the binary classification model is able to detect antifungal drugs (tavaborole, an antifungal protein synthesis inhibitor) with a mechanism of action which was not present in the training data set. This is a highly desirable property of any phenotypic screening assay, as it shows that the assay allows for the identification of drugs with novel mechanisms of action. Lastly, the proteomic effect of different types of drugs on mammalian cells was explored by using the HeLa cancer cell line. It was shown that the presented proteomic profiling approach can easily detect several types of drug-induced stresses in HeLa cells, in particular corticosteroids and tubulin (de)polymerization inhibitors, but is less suitable for distinguishing other types of drug classes (neurotransmitter antagonists, statins, opioids). Additionally, the application of the assay on HeLa cells demonstrated the ability to detect different types of stresses, such as the cells’ proteomic response to UV exposure or heat shocks. These results pave the way for possible distinction between apoptosis and necrosis pathways in HeLa cells using the presented MALDI-TOF MS based method. 
\nTo conclude, a high-throughput compatible, label free, MALDI-TOF mass spectrometry-based screening assay is described in this thesis, which measures sub-lethal drug effects on the cellular proteome in a phenotypic and pharmacological relevant setting. The method was found suitable for whole-cell screening of small libraries of drugs, and showed the ability to distinguish different types of stresses elicited on multiple types of cell cultures. The potential to find new, weakly active drugs within a known mechanism of action, as well as the ability to detect sub-lethal drug responses with new mechanisms of action for which the model was not trained, was demonstrated. The characteristic to identify novel mechanisms of action in a cell-based screen can be exploited to solve the most pressing issues in drug discovery today. In addition, mechanistic information of the drugs activity can be used as a starting point for further target elucidation or to prioritize drug screening hits. The studies performed in this thesis have resulted in a solid foundation for further research that expand the capabilities of the MALDI-TOF MS-based assay in a broad range of phenotypic profiling applications in the drug discovery field.

Image Classification (IC) is most prominent among other Artificial Intelligence (AI) domains. Mainly, IC participates rigorously for the development of society in a variety of application areas such as finance, marketing, health, industrial automation, education, and safety and security. Typically, an IC model takes image input data and tunes itself as per the required application task and classify accordingly. Among the various categories of images, color image category is better due to the capability of capturing more details, which are essential for classification purpose. However, the modern world demands Realtime or online image classification, which involves Imagery Streams. The highly likely uncertainty in Imagery Streams is due to non-stationary environment, for example, certain features or class boundaries which are valid at one-time step are not adequate for another time step. These uncertainties in Imagery Streams have deleterious effects on IC models, which causes performance degradation in terms of accuracy or make IC models, not in further use. Therefore, to overcome these issues, IC models need to adapt to changes caused by uncertainties in Imagery Streams. This paper focuses on the understanding the possible scenarios of such uncertainties in Color Imagery Streams, investigates the deleterious effects due to changes in Color Imagery Streams and provides the possible mitigation approach to overcome the issues in IC models. The contribution of this research is the first step towards an adaptive model development to mitigate the deleterious effects of uncertainty in Color Imagery Streams. This model will benefit many application areas and will directly contribute to the daily life of a society.

Recently, number of medical X-ray images being generated is increasing rapidly due to the advancements in radiological equipment in medical centres. Medical X-ray image classification techniques are needed for effective decision making in the healthcare sector. Since the traditional image classification models are ineffective to accomplish maximum X-ray image classification performance, deep learning (DL) models have emerged. In this study, an Arithmetic Optimization Algorithm with Deep Learning-Based Medical X-Ray Image Classification (AOADL-MXIC) model has been developed. The proposed AOADL-MXIC model investigates the available X-ray images for the identification of diseases. Initially, the AOADL-MXIC model executes the pre-processing step using the Gabor filtering (GF) technique to eliminate the presence of noise. In the next level, the Capsule Network (CapsNet) model is utilized to derive feature vectors from the input X-ray images. Furthermore, for optimizing the hyperparameters related to the CapsNet approach, the AOA is exploited. Finally, the bidirectional gated recurrent unit (BiGRU) model is employed for the classification of medical X-ray images. The experimental result analysis of the AOADL-MXIC technique on a set of medical images stated the promising performance over the other models.KeywordsX-ray imagesArithmetic optimization algorithmDeep learningFeature extractionHyperparameter tuning

We introduce the Plant Disease Detection Platform (PDDP) that allows users to send photos of sick plant leaves or textual descriptions of their appearance to obtain information about an infection that hit the vegetation and treatment tips. The backend of the platform in terms of deep learning includes image classification and text similarity models. The image classification model has two parts: feature extractor and classifier. The feature extractor was trained using the triplet loss function along with transfer learning when the weights of the network are initialized from the MobileNetV2 pretrained on the ImageNet dataset. The classifier is a simple multilayer perceptron. The test on 100 random plant images from the Internet shows 98% of the classification accuracy. We did the post-training static quantization in order to reduce the overall model size and increase the inference performance. The final model has a size of 7 Mb and works 5 times faster than the initial model without significant loss of accuracy. The text similarity model is a BERT-based transformer for obtaining vector representation of input texts for further similarity calculation between user requests and disease descriptions on the PDDP.

Aiming at the problem of image classification with insignificant morphological structural features, strong target correlation, and low signal-to-noise ratio, combined with prior feature knowledge embedding, a deep learning method based on ResNet and Radial Basis Probabilistic Neural Network (RBPNN) is proposed model. Taking ResNet50 as a visual modeling network, it uses feature pyramid and self-attention mechanism to extract appearance and semantic features of images at multiple scales, and associate and enhance local and global features. Taking into account the diversity of category features, channel cosine similarity attention and dynamic C-means clustering algorithms are used to select representative sample features in different category of sample subsets to implicitly express prior category feature knowledge, and use them as the kernel centers of radial basis probability neurons (RBPN) to realize the embedding of diverse prior feature knowledge. In the RBPNN pattern aggregation layer, the outputs of RBPN are selectively summed according to the category of the kernel center, that is, the subcategory features are combined into category features, and finally the image classification is implemented based on Softmax. The functional module of the proposed method is designed specifically for image characteristics, which can highlight the significance of local and structural features of the image, form a non-convex decision-making area, and reduce the requirements for the completeness of the sample set. Applying the proposed method to medical image classification, experiments were conducted based on the brain tumor MRI image classification public dataset and the actual cardiac ultrasound image dataset, and the accuracy rate reached 85.82% and 83.92% respectively. Compared with the three mainstream image classification models, the performance indicators of this method have been significantly improved.