Classifier Construction Research Articles

Lung adenocarcinoma identification based on hybrid feature selections and attentional convolutional neural networks.

Lung adenocarcinoma, a chronic non-small cell lung cancer, needs to be detected early. Tumor gene expression data analysis is effective for early detection, yet its challenges lie in a small sample size, high dimensionality, and multi-noise characteristics. In this study, we propose a lung adenocarcinoma convolutional neural network (LATCNN), a deep learning model tailored for accurate lung adenocarcinoma prediction and identification of key genes. During the feature selection stage, we introduce a hybrid algorithm. Initially, the fast correlation-based filter (FCBF) algorithm swiftly filters out irrelevant features, followed by applying the k-means-synthetic minority over-sampling technique (k-means-SMOTE) method to address category imbalance. Subsequently, we enhance the particle swarm optimization (PSO) algorithm by incorporating fast-decay dynamic inertia weights and utilizing the classification and regression tree (CART) as the fitness function for the second stage of feature selection, aiming to further eliminate redundant features. In the classifier construction stage, we present an attention convolutional neural network (atCNN) that incorporates an attention mechanism. This improved model conducts feature selection post lung adenocarcinoma gene expression data analysis for classification and prediction. The results show that LATCNN effectively reduces the feature dimensions and accurately identifies 12 key genes with accuracy, recall, F1 score, and MCC of 99.70%, 99.33%, 99.98%, and 98.67%, respectively. These performance metrics surpass those of other comparative models, highlighting the significance of this research for advancing lung adenocarcinoma treatment.

Differentiation and risk stratification of basal cell carcinoma with deep learning on histopathologic images and measuring nuclei and tumor microenvironment features.

Nuclear pleomorphism and tumor microenvironment (TME) play a critical role in cancer development and progression. Identifying most predictive nuclei and TME features of basal cell carcinoma (BCC) may provide insights into which characteristics pathologists can use to distinguish and stratify this entity. To develop an automated workflow based on nuclei and TME features from basaloid cell tumor regions to differentiate BCC from trichoepithelioma (TE) and stratify BCC into high-risk (HR) and low-risk (LR) subtypes, and to identify the nuclear and TME characteristics profile of different basaloid cell tumors. The deep learning systems were trained on 161 H&E -stained sections which contained 51 sections of HR-BCC, 50 sections of LR-BCC and 60 sections of TE from one institution (D1), and externally and independently validated on D2 (46 sections) and D3 (76 sections), from 2015 to 2022. 60%, 20% and 20% of D1 data were randomly splitted for training, validation and testing, respectively. The framework comprised four stages: tumor regions identification by multi-head self-attention (MSA) U-Net, nuclei segmentation by HoVer-Net, quantitative feature by handcrafted extraction, and differentiation and risk stratification classifier construction. Pixel accuracy, precision, recall, dice score, intersection over union (IoU) and area under the curve (AUC) were used to evaluate the performance of tumor segmentation model and classifiers. MSA-U-Net model detected tumor regions with 0.910 precision, 0.869 recall, 0.889 dice score and 0.800 IoU. The differentiation classifier achieved 0.977±0.0159, 0.955±0.0181, 0.885±0.0237 AUC in D1, D2 and D3, respectively. The most discriminative features between BCC and TE contained Homogeneity, Elongation, T-T_meanEdgeLength, T-T_Nsubgraph, S-T_HarmonicCentrality, S-S_Degrees. The risk stratification model can well predict HR-BCC and LR-BCC with 0.920±0.0579, 0.839±0.0176, 0.825±0.0153 AUC in D1, D2 and D3, respectively. The most discriminative features between HR-BCC and LR-BCC comprised IntensityMin, Solidity, T-T_minEdgeLength, T-T_Coreness, T-T_Degrees, T-T_Betweenness, S-T_Degrees. This framework hold potential for future use as a second opinion helping inform diagnosis of BCC, and identify nuclei and TME features related with malignancy and tumor risk stratification.

Prediction of lymph node metastasis of lung squamous cell carcinoma by machine learning algorithm classifiers.

Lymph node metastasis (LNM) is an essential factor affecting the prognosis of patients with lung squamous cell carcinoma (LUSC), as well as a critical consideration for the choice of treatment strategy. Exploring effective methods for predicting LNM in LUSC may benefit clinical decision making. We used data collected from the Surveillance, Epidemiology, and End Results (SEER) database to develop machine learning algorithm classifiers, including boosted trees (BTs), based on the primary clinical parameters of patients to predict LNM in LUSC. Training on a large-sample training cohort (n = 8,063) allowed for the construction of several concise classifiers for LNM prediction in LUSC, which were then validated using test and in-house cohorts (n = 2,017 and 57, respectively). The six classifiers established in this research enabled distinction between patients with and without LNM. Among these classifiers, the BT classifier was the top performer, with accuracy, F1 scores, precision, recall, sensitivity, and specificity values of 0.654, 0.621, 0.654, 0.592, 0.592, and 0.711, respectively. The precision recall (PR) and receiver operating characteristic (ROC) (with area under the curve = 0.714) curves also supported this result, which was validated by the in-house cohort. Notably, the tumor stage was a critical factor in determining LNM in patients with LUSC. The use of classifiers, especially the BT classifier, may serve as a useful tool for improving clinical precision and individualized treatment of patients with LUSC.

Developing a classifier to screen for mild cognitive impairment and Alzheimer’s disease using machine learning approaches

AbstractBackgroundMachine learning classifier is a powerful tool that has been recently applied to datasets of Alzheimer’s disease (AD). The majority of approaches developed to classify people with mild cognitive impairment (MCI) and Alzheimer’s disease (AD) from cognitively normal (CN) people have made use of expensive or invasive measurements, discouraging the clinical use of these tools. In the current study, we aimed to develop a classifier to screen for MCI and AD using data that can be routinely collected at low cost (e.g. blood pressure, ApoE genotype).MethodDatabase of the Australian Imaging Biomarkers and Lifestyle Study of Ageing (AIBL) was used (n = 14553 data entries). Dataset was split into two parts: training (80%) and testing (20%). A decision tree was first employed to screen for factors (i.e. clinical parameters) that can effectively discriminate CN, MCI, and AD (Step 1). To further refine factor selection, the association between the factors (from screening) and health status (CN, MCI, AD) were assessed using linear and logistic regression methods. Factors that were rejected by the hypothesis test (i.e. non‐obvious relationship with MCI or AD) and those require high cost for collection were removed (Step 2). The remaining factors were used for the construction of the classifier, where decision tree, Gaussian Naïve Bayes and second‐order polynomial approaches were employed.ResultsThe selected factors for classifier construction were ApoE genotype (p<0.001), medical history of stroke (p<0.001 AD, p = 0.223 MCI), diabetes (p<0.001) and depression (p<0.001), and systolic blood pressure (p<0.001). Three models were used to construct the classifier, and the differential accuracy ranges between 74.1 – 84.2% for CN, 25.3 – 50.2% for MCI, and 47.8 – 73.4% for AD (Table 1). Accuracy from second‐order polynomial model was lower than that from the decision tree and Gaussian Naïve Bayes model.ConclusionThe classifier developed in the current study is yet to be optimized, especially for MCI. In future study, the performance of our classifier will be evaluated using other AD datasets for generalizability. This classifier can potentially be used by the clinicians as an MCI/AD screening tool for at‐risk population prior to requesting costly diagnostic tests.

Combined Vorinostat and Chloroquine Inhibit Sodium-Iodide Symporter Endocytosis and Enhance Radionuclide Uptake In Vivo.

Patients with aggressive thyroid cancer are frequently failed by the central therapy of ablative radioiodide (RAI) uptake, due to reduced plasma membrane (PM) localization of the sodium/iodide symporter (NIS). We aimed to understand how NIS is endocytosed away from the PM of human thyroid cancer cells, and whether this was druggable in vivo. Informed by analysis of endocytic gene expression in patients with aggressive thyroid cancer, we used mutagenesis, NanoBiT interaction assays, cell surface biotinylation assays, RAI uptake, and NanoBRET to understand the mechanisms of NIS endocytosis in transformed cell lines and patient-derived human primary thyroid cells. Systemic drug responses were monitored via 99mTc pertechnetate gamma counting and gene expression in BALB/c mice. We identified an acidic dipeptide within the NIS C-terminus that mediates binding to the σ2 subunit of the Adaptor Protein 2 (AP2) heterotetramer. We discovered that the FDA-approved drug chloroquine (CQ) modulates NIS accumulation at the PM in a functional manner that is AP2 dependent. In vivo, CQ treatment of BALB/c mice significantly enhanced thyroidal uptake of 99mTc pertechnetate in combination with the histone deacetylase (HDAC) inhibitor vorinostat/SAHA, accompanied by increased thyroidal NIS mRNA. Bioinformatic analyses validated the clinical relevance of AP2 genes with disease-free survival in RAI-treated DTC, enabling construction of an AP2 gene-related risk score classifier for predicting recurrence. NIS internalization is specifically druggable in vivo. Our data, therefore, provide new translatable potential for improving RAI therapy using FDA-approved drugs in patients with aggressive thyroid cancer. See related commentary by Lechner and Brent, p. 1220.

Construction and testing of a risk prediction classifier for cardia carcinoma.

This research aimed to construct a prediction model for stages II and III cardia carcinoma (CC), and provide an effective preoperative evaluation tool for clinicians. CC mRNA expression matrix was obtained from Gene Expression Omnibus and The Cancer Genome Atlas databases. Non-negative matrix factorization was used to cluster data to obtain subgroup information, and weighted gene co-expression network analysis was used to uncover key modules linked to different subgroups. Gene-set enrichment analysis analyzed biological pathways of different subgroups. The related pathways of multiple modules were scrutinized with Kyoto Encyclopedia of Genes and Genomes. Key modules were manually annotated to screen CC-related genes. Subsequently, quantitative real-time polymerase chain reaction assessed CC-related gene expression in fresh tissues and paraffin samples, and Pearson correlation analysis was performed. A classification model was constructed and the predictive ability was evaluated by the receiver operating characteristic curve. CC patients had four subgroups that were associated with brown, turquoise, red, and black modules, respectively. The CC-related modules were mainly associated with abnormal cell metabolism and inflammatory immune pathways. Then, 76 CC-elated genes were identified. Pearson correlation analysis presented that THBS4, COL14A1, DPYSL3, FGF7, and SVIL levels were relatively stable in fresh and paraffin tissues. The area under the curve of 5-gene combined prediction for staging was 0.8571, indicating good prediction ability. The staging classifier for CC based on THBS4, COL14A1, DPYSL3, FGF7, and SVIL has a good predictive effect, which may provide effective guidance for whether CC patients need emergency surgery.

Construction and validation of a gene expression classifier to predict immunotherapy response in primary triple-negative breast cancer

BackgroundImmune checkpoint inhibitors (ICI) improve clinical outcomes in triple-negative breast cancer (TNBC) patients. However, a subset of patients does not respond to treatment. Biomarkers that show ICI predictive potential in other solid tumors, such as levels of PD-L1 and the tumor mutational burden, among others, show a modest predictive performance in patients with TNBC.MethodsWe built machine learning models based on pre-ICI treatment gene expression profiles to construct gene expression classifiers to identify primary TNBC ICI-responder patients. This study involved 188 ICI-naïve and 721 specimens treated with ICI plus chemotherapy, including TNBC tumors, HR+/HER2− breast tumors, and other solid non-breast tumors.ResultsThe 37-gene TNBC ICI predictive (TNBC-ICI) classifier performs well in predicting pathological complete response (pCR) to ICI plus chemotherapy on an independent TNBC validation cohort (AUC = 0.86). The TNBC-ICI classifier shows better performance than other molecular signatures, including PD-1 (PDCD1) and PD-L1 (CD274) gene expression (AUC = 0.67). Integrating TNBC-ICI with molecular signatures does not improve the efficiency of the classifier (AUC = 0.75). TNBC-ICI displays a modest accuracy in predicting ICI response in two different cohorts of patients with HR + /HER2- breast cancer (AUC = 0.72 to pembrolizumab and AUC = 0.75 to durvalumab). Evaluation of six cohorts of patients with non-breast solid tumors treated with ICI plus chemotherapy shows overall poor performance (median AUC = 0.67).ConclusionTNBC-ICI predicts pCR to ICI plus chemotherapy in patients with primary TNBC. The study provides a guide to implementing the TNBC-ICI classifier in clinical studies. Further validations will consolidate a novel predictive panel to improve the treatment decision-making for patients with TNBC.

Thermal desorption direct analysis in real-time high-resolution mass spectrometry and machine learning allow the rapid authentication of ground black pepper and dried oregano: A proof-of-concept study.

Thermal desorption direct analysis in real-time high-resolution mass spectrometry (TD-DART-HRMS) approaches have gained popularity for fast screening of a variety of samples. With rapid volatilization of the sample at increasing temperatures outside the mass spectrometer, this technique can provide a direct readout of the sample content with no sample preparation. In this study, TD-DART-HRMS's utility for establishing spice authenticity was examined. To this aim, we directly analyzed authentic (typical) and adulterated (atypical) samples of ground black pepper and dried oregano in positive and negative ion modes. We analyzed a set of authentic ground black pepper samples (n = 14) originating from Brazil, Sri Lanka, Madagascar, Ecuador, Vietnam, Costa Rica, Indonesia, Cambodia, and adulterated samples (n = 25) consisting of mixtures of ground black pepper with this spice's nonfunctional by-products (pinheads or spent) or with different exogenous materials (olive kernel, green lentils, black mustard seeds, red beans, gypsum plaster, garlic, papaya seeds, chili, green aniseed, or coriander seeds). TD-DART-HRMS facilitated the capture of informative fingerprinting of authentic dried oregano (n = 12) originating from Albania, Turkey, and Italy and those spiked (n = 12) with increasing percentages of olive leaves, sumac, strawberry tree leaves, myrtle, and rock rose. A predictive LASSO classifier was built, after merging by low-level data fusion, the positive and negative datasets for ground black pepper. Fusing multimodal data allowed retrieval of more comprehensive information from both datasets. The resultant classifier achieved on the withheld test set accuracy, sensitivity, and specificity of 100%, 75%, and 90%, respectively. On the contrary, the sole TD-(+)DART-HRMS spectra of the oregano samples allowed construction of a LASSO classifier that predicted the adulteration of the oregano with excellent statistical indicators. This classifier achieved, on the withheld test set, 100% each for accuracy, sensitivity, and specificity.

Construction of subtype classifiers and validation of a prognostic risk model based on hypoxia-associated lncRNAs for lung adenocarcinoma.

Studies have shown that long non-coding RNAs (lncRNAs) are found to be hypoxia-regulated lncRNAs in cancer. Lung adenocarcinoma (LUAD) is the leading cause of cancer death worldwide, and despite early surgical removal, has a poor prognosis and a high recurrence rate. Thus, we aimed to identify subtype classifiers and construct a prognostic risk model using hypoxia-associated long noncoding RNAs (hypolncRNAs) for LUAD. Clinical data of LUAD samples with prognosis information obtained from the Gene Expression Omnibus (GEO), acted as validation dataset, and The Cancer Genome Atlas (TCGA) databases, served as training dataset, were used to screen hypolncRNAs in each dataset by univariate Cox regression analysis; the intersection set was used for subsequent analyses. Unsupervised clustering analysis was performed based on the expression of hypolncRNAs using the 'ConsensuClusterPlus' package. The tumor microenvironment (TME) was compared between LUAD subgroups by analyzing the expression of immune cell infiltration, immune components, stromal components, immune checkpoints, and chemokine secretion. To identify robust prognostically associated hypolncRNAs and construct a risk score model, multivariate Cox regression analysis was performed. A total of 14 hypolncRNAs were identified. Based on the expression of these hypolncRNAs, patients with LUAD were classified into three hypolncRNA-regulated subtypes. The three subtypes differed significantly in immune cell infiltration, stromal score, specific immune checkpoints, and secretion of chemokines and their receptors. The Tumor Immune Dysfunction and Exclusion (TIDE) scores and myeloid-derived suppressor cell (MDSC) scores were also found to differ significantly among the three hypolncRNA-regulated subtypes. Four of the 14 hypolncRNAs were used to construct a signature to distinguish the overall survival (OS) in TCGA dataset (P<0.0001) and GEO dataset (P=0.0032) and sensitivity to targeted drugs in patients at different risks of LUAD. We characterized three regulatory subtypes of hypolncRNAs with different TMEs. We developed a signature based on hypolncRNAs, contributing to the development of personalized therapy and representing a new potential therapeutic target for LUAD.

Sex-dependent performance differences in curvilinear aiming arm movements in octogenarians

In an aging society, it is necessary to detect the cognitive decline of individuals at an early stage using simple measurement methods. This makes early health care possible for those affected. The aim of the study was to develop a classifier for cognitive state in older adults with and without mild cognitive impairment (MCI) based on kinematic parameters of linear and curvilinear aiming arm movements. In a group of 224 older adults over 80 years of age (cognitively healthy and MCI), the movement duration and intersegment intervals of linear and curvilinear arm movements of 20 cm were recorded. Movement duration was significantly longer in the curvilinear condition than in the straight movement, and MCI participants required significantly more time than cognitively healthy participants. Post-hoc analysis on the fluidity of movement in the curvilinear condition showed that MCI men had significantly longer inter-segmental intervals than non-MCI men. No difference was found in women. Based on the inter-segmental intervals, a simple classifier could be developed that correctly classified 63% of the men. In summary, aiming arm movements are only conditionally suitable as a classifier for cognitive states. For the construction of an ideal classifier, age-related degeneration of cortical and subcortical motor areas should be considered.

Kinect-ing the Dots: Using Motion-Capture Technology to Distinguish Sign Language Linguistic From Gestural Expressions.

Just as vocalization proceeds in a continuous stream in speech, so too do movements of the hands, face, and body in sign languages. Here, we use motion-capture technology to distinguish lexical signs in sign language from other common types of expression in the signing stream. One type of expression is constructed action, the enactment of (aspects of) referents and events by (parts of) the body. Another is classifier constructions, the manual representation of analogue and gradient motions and locations simultaneously with specified referent morphemes. The term signing is commonly used for all of these, but we show that not all visual signals in sign languages are of the same type. In this study of Israeli Sign Language, we use motion capture to show that the motion of lexical signs differs significantly along several kinematic parameters from that of the two other modes of expression: constructed action and the classifier forms. In so doing, we show how motion-capture technology can help to define the universal linguistic category "word," and to distinguish it from the expressive gestural elements that are commonly found across sign languages.

Characterization of stemness features and construction of a stemness subtype classifier to predict survival and treatment responses in lung squamous cell carcinoma

BackgroundCancer stemness has been proven to affect tumorigenesis, metastasis, and drug resistance in various cancers, including lung squamous cell carcinoma (LUSC). We intended to develop a clinically applicable stemness subtype classifier that could assist physicians in predicting patient prognosis and treatment response.MethodsThis study collected RNA-seq data from TCGA and GEO databases to calculate transcriptional stemness indices (mRNAsi) using the one-class logistic regression machine learning algorithm. Unsupervised consensus clustering was conducted to identify a stemness-based classification. Immune infiltration analysis (ESTIMATE and ssGSEA algorithms) methods were used to investigate the immune infiltration status of different subtypes. Tumor Immune Dysfunction and Exclusion (TIDE) and Immunophenotype Score (IPS) were used to evaluate the immunotherapy response. The pRRophetic algorithm was used to estimate the efficiency of chemotherapeutic and targeted agents. Two machine learning algorithms (LASSO and RF) and multivariate logistic regression analysis were performed to construct a novel stemness-related classifier.ResultsWe observed that patients in the high-mRNAsi group had a better prognosis than those in the low-mRNAsi group. Next, we identified 190 stemness-related differentially expressed genes (DEGs) that could categorize LUSC patients into two stemness subtypes. Patients in the stemness subtype B group with higher mRNAsi scores exhibited better overall survival (OS) than those in the stemness subtype A group. Immunotherapy prediction demonstrated that stemness subtype A has a better response to immune checkpoint inhibitors (ICIs). Furthermore, the drug response prediction indicated that stemness subtype A had a better response to chemotherapy but was more resistant to epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs). Finally, we constructed a nine-gene-based classifier to predict patients’ stemness subtype and validated it in independent GEO validation sets. The expression levels of these genes were also validated in clinical tumor specimens.ConclusionThe stemness-related classifier could serve as a potential prognostic and treatment predictor and assist physicians in selecting effective treatment strategies for patients with LUSC in clinical practice.

Random Forest Classifier of Heme Proteins Using Porphyrin Distortions and Axial Ligands of Heme

We achieved success in the construction of a random forest classifier that is capable of accurately classifying heme proteins for oxygen storage/transport function and redox function, with a high level of precision (>83%), by utilizing heme porphyrin distortions and axial ligands of heme as features. Our findings suggest that the heme distortions and the axial ligands are significantly correlated with oxidoreductase functions and oxygen storage/transport functions, respectively, all of which are commonly observed in heme proteins. We constructed a random forest classifier using heme porphyrin distortions and axial ligands as features to successfully classify and predict two distinct classes of heme protein functions with remarkable accuracy. Our findings revealed that heme porphyrin distortion played a significant role in the classification of oxidoreductases, while axial ligand was found to be crucial for the classification of oxygen-binding proteins.

The RNA-Binding Proteins OAS1, ZFP36L2, and DHX58 Are Involved in the Regulation of CD44 mRNA Splicing in Colorectal Cancer Cells.

Regulation of alternative splicing is carried out by RNA-binding proteins. Each alternative splicing event is controlled by several RNA-binding proteins, which in combination create the distribution of alternative splicing products in a given cell type. Transmembrane protein CD44 plays an important role at various stages of the metastatic cascade and is considered as a promising molecule for the therapy of tumor diseases and the construction of prognostic classifiers. However, the functions of specific isoforms of this protein may differ significantly. In this work, we performed a bioinformatic search of RNA-binding proteins that can determine the expression of clinically significant isoforms 3 and 4 of CD44 protein. The analysis revealed five RNA-binding proteins, three of which (OAS1, ZFP36L2, and DHX58) are shown for the first time as potential regulators of the studied process.

Subphenotyping heterogeneous patients with chronic critical illness to guide individualised fluid balance treatment using machine learning: a retrospective cohort study.

The great heterogeneity of patients with chronic critical illness (CCI) leads to difficulty for intensive care unit (ICU) management. Identifying subphenotypes could assist in individualized care, which has not yet been explored. In this study, we aim to identify the subphenotypes of patients with CCI and reveal the heterogeneous treatment effect of fluid balance for them. In this retrospective study, we defined CCI as an ICU length of stay over 14 days and coexists with persistent organ dysfunction (cardiovascular Sequential Organ Failure Assessment (SOFA) score ≥1 or score in any other organ system ≥2) at Day 14. Data from five electronic healthcare record datasets covering geographically distinct populations (the US, Europe, and China) were studied. These five datasets include (1) subset of Derivation (MIMIC-IV v1.0, US) cohort (2008-2019); (2) subset Derivation (MIMIC-III v1.4 'CareVue', US) cohort (2001-2008); (3) Validation I (eICU-CRD, US) cohort (2014-2015); (4) Validation II (AmsterdamUMCdb/AUMC, Euro) cohort (2003-2016); (5) Validation III (Jinling, CN) cohort (2017-2021). Patients who meet the criteria of CCI in their first ICU admission period were included in this study. Patients with age over 89 or under 18 years old were excluded. Three unsupervised clustering algorithms were employed independently for phenotypes derivation and validation. Extreme Gradient Boosting (XGBoost) was used for phenotype classifier construction. A parametric G-formula model was applied to estimate the cumulative risk under different daily fluid management strategies in different subphenotypes of ICU mortality. We identified four subphenotypes as Phenotype A, B, C, and D in a total of 8145 patients from three countries. Phenotype A is the mildest and youngest subgroup; Phenotype B is the most common group, of whom patients showed the oldest age, significant acid-base abnormality, and low white blood cell count; Patients with Phenotype C have hypernatremia, hyperchloremia, and hypercatabolic status; and in Phenotype D, patients accompany with the most severe multiple organ failure. An easy-to-use classifier showed good effectiveness. Phenotype characteristics showed robustness across all cohorts. The beneficial fluid balance threshold intervals of subphenotypes were different. We identified four novel phenotypes that revealed the different patterns and significant heterogeneous treatment effects of fluid therapy within patients with CCI. A prospective study is needed to validate our findings, which could inform clinical practice and guide future research on individualized care. This study was funded by 333 High Level Talents Training Project of Jiangsu Province (BRA2019011), General Program of Medical Research from the Jiangsu Commission of Health (M2020052), and Key Research and Development Program of Jiangsu Province (BE2022823).

중국어 동량사구문의 형식과 의미의 일반화 연구

중국어 동량사구문의 형식과 의미의 일반화 연구

A Textual Backdoor Defense Method Based on Deep Feature Classification

Natural language processing (NLP) models based on deep neural networks (DNNs) are vulnerable to backdoor attacks. Existing backdoor defense methods have limited effectiveness and coverage scenarios. We propose a textual backdoor defense method based on deep feature classification. The method includes deep feature extraction and classifier construction. The method exploits the distinguishability of deep features of poisoned data and benign data. Backdoor defense is implemented in both offline and online scenarios. We conducted defense experiments on two datasets and two models for a variety of backdoor attacks. The experimental results demonstrate the effectiveness of this defense approach and outperform the baseline defense method.

Design of Efficient Speech Emotion Recognition Based on Multi Task Learning

Speech emotion recognition technology includes feature extraction and classifier construction. However, the recognition efficiency is reduced due to noise interference and gender differences. To solve this problem, this paper used two multi-task learning models based on adversarial multi-task learning(ASP-MTL). The first model took emotion recognition as the main task and noise recognition as the auxiliary task, and removed the noise part identified by the auxiliary task. After identifying the non-noise part, the second model was constructed. The second model took emotion recognition as the main task and gender classification as the auxiliary task. These two multi-task learning models can not only can use shared information to learn the relationship between different tasks, but also can identify specific tasks. This paper used Audio/Visual Emotion Challenge (AVEC) database and AFEW6.0 database,which were recorded in the field environment. Considering the problem of data imbalance between datasets, the data balance operation was carried out on the data sets in the process of data preprocessing. The paper shows an increase of around 10% in terms of accuracy and F1 score with the recent works using AVEC database and AFEW6.0 datasets, which proved that this paper has made a great progress in SER.

Development of a multi-channel classifier of rail line states

The article deals with the construction of a three-channel invariant classifier that has the properties of classifying the states of rail lines into a set of classes that are invariant to changes in the longitudinal resistance of the rail line and the transverse conductivity of the insulation of the ballast material. Invariance is achieved taking into account the change in the transverse conductivity of the insulation and the longitudinal resistance of the rail line while compiling systems of equations of state for rail lines, which are the decisive functions of the classifier. The article shows that the three-channel method allows for the correct recognition of all three classes of rail line states by three decision functions with arguments - voltages and currents at the input and output of the rail line. The block diagram of the operation algorithm of the three-channel classifier of the states of the rail lines allows to form the recognition process and the majority classification depending on the states of the channels. The feasibility of the algorithm is confirmed by simulation studies on a mathematical model and graphical results.

The Structure of Multiple Classifier Constructions and Some Related Theoretical Issues

This paper addresses multiple classifier constructions (MCCs) in terms of internal structure and derivation. Based on a summary of the major features of MCCs, the paper proposes that MCCs are composed of DP and NumP. The combination of the two numeral-classifiers denotes a complete structure, which presents the pattern of double numeral-classifiers. DP serves as the head while NumP serves as the complement. There is no coordination but topic-comment relation between the embedded DP and NumP. The embedded DP c-commands NumP. NumP moves across the board to the specifier position of the external DP to give rise to various types of MCCs. Movement of the classifier of MCCs is triggered by numeral gapping. Based on this, it is argued that all MCCs must be of this nature in terms of the underlying structure. This contention can provide a reasonable solution to many long-standing puzzles related to MCCs.