Unsupervised Module Research Articles

Synergizing Unsupervised and Supervised Learning: A Hybrid Approach for Accurate Natural Language Task Modeling

While supervised learning models have shown remarkable performance in various natural language processing (NLP) tasks, their success heavily relies on the availability of large-scale labeled datasets, which can be costly and time-consuming to obtain. Conversely, unsupervised learning techniques can leverage abundant unlabeled text data to learn rich representations, but they do not directly optimize for specific NLP tasks. This paper presents a novel hybrid approach that synergizes unsupervised and supervised learning to improve the accuracy of NLP task modeling. While supervised models excel at specific tasks, they rely on large labeled datasets. Unsupervised techniques can learn rich representations from abundant unlabeled text but don't directly optimize for tasks. Our methodology integrates an unsupervised module that learns representations from unlabeled corpora (e.g., language models, word embeddings) and a supervised module that leverages these representations to enhance task-specific models [4]. We evaluate our approach on text classification and named entity recognition (NER), demonstrating consistent performance gains over supervised baselines. For text classification, contextual word embeddings from a language model pretrain a recurrent or transformer-based classifier. For NER, word embeddings initialize a BiLSTM sequence labeler. By synergizing techniques, our hybrid approach achieves SOTA results on benchmark datasets, paving the way for more data-efficient and robust NLP systems.

A clinical consensus-compliant deep learning approach to quantitatively evaluate human in vitro fertilization early embryonic development with optical microscope images

The selection of embryos is a key for the success of in vitro fertilization (IVF). However, automatic quality assessment on human IVF embryos with optical microscope images is still challenging. In this study, we developed a clinical consensus-compliant deep learning approach, named Esava (Embryo Segmentation and Viability Assessment), to quantitatively evaluate the development of IVF embryos using optical microscope images. In total 551 optical microscope images of human IVF embryos of day-2 to day-3 were collected, preprocessed, and annotated. Using the Faster R-CNN model as baseline, our Esava model was constructed, refined, trained, and validated for precise and robust blastomere detection. A novel algorithm Crowd-NMS was proposed and employed in Esava to enhance the object detection and to precisely quantify the embryonic cells and their size uniformity. Additionally, an innovative GrabCut-based unsupervised module was integrated for the segmentation of blastomeres and embryos. Independently tested on 94 embryo images for blastomere detection, Esava obtained the high rates of 0.9940, 0.9121, and 0.9531 for precision, recall, and mAP respectively, and gained significant advances compared with previous computational methods. Intraclass correlation coefficients indicated the consistency between Esava and three experienced embryologists. Another test on 51 extra images demonstrated that Esava surpassed other tools significantly, achieving the highest average precision 0.9025. Moreover, it also accurately identified the borders of blastomeres with mIoU over 0.88 on the independent testing dataset. Esava is compliant with the Istanbul clinical consensus and compatible to senior embryologists. Taken together, Esava improves the accuracy and efficiency of embryonic development assessment with optical microscope images.

Few-shot classification with multisemantic information fusion network

Metric-based methods aim to predict class labels by computing the similarity between samples using distance functions, which is the mainstream approach to few-shot learning. However, the limited representational space of feature vectors and appearance variations among congenetic samples still present challenges. We propose a Multisemantic Information Fusion Network (MIFN) to address these problems. A Lower-level Feature generator (LF-generator), which is an unsupervised module, adaptively activates high-response regions of objects to introduce discriminative semantic details. Meanwhile, a Higher-level Feature extractor (HF-extractor) learns global semantic information with human cognition to minimise the impact of appearance variations. We integrate the coarse outputs of these two modules, which complement each other to jointly promote more precise predictions. Furthermore, considering the importance of prototypes, we redefine the sampling strategy of the triplet loss and utilise it as an auxiliary loss to sharpen the decision boundary at the prototype level, facilitating subsequent classification. Our experimental results demonstrate the competitiveness of our approach in both general few-shot classification (mini-ImageNet and tiered-ImageNet) and cross-domain problems (CUB, Caltech-101, Stanford-dogs, and Stanford-cars) with minimal bells and whistles.

Learning zero-shot dense light field reconstruction from heterogeneous imaging

Due to the limited sensor size, light field (LF) imaging usually suffers from an essential trade-off between its spatial and angular resolutions. Recently, some approaches have been proposed to use hybrid lens imaging to enhance spatial resolution, but LF angular reconstruction has not been involved. In this paper, we propose a zero-shot learning-based method for reconstructing a dense LF image from a heterogeneous imaging system that integrates a 2D digital camera and a commercial LF camera. Our main insight is that the high-resolution information of the 2D digital image can assist in enhancing the spatial resolution of the LF image, and further, the spatially enhanced LF image can guide the angular reconstruction of the 2D digital image. To avoid ambiguity, it is pointed out that this paper focuses on the latter, i.e., enhancing the angular resolution of the 2D image. Specifically, we first construct two unsupervised modules to learn the geometric property of the acquired LF image and the correlation between the LF image and 2D image, so as to perform the initial reconstruction. Afterwards, an LF detail refiner is constructed to improve the reconstruction quality, and a cycle consistency constraint is designed to realize unsupervised training. Consequently, we can learn the dense LF reconstruction from the heterogeneous imaging data itself without extra training samples. Experimental results show that the proposed method can faithfully reconstruct high-quality densely-sampled LF images, and surpasses some state-of-the-art methods in both quantitative and qualitative evaluations. Furthermore, the proposed method can be combined with existing LF view stitching methods to further enhance the field of view of the reconstructed LF image. Our code will be released at https://github.com/YeyaoChen/ZLFSR.

Generative Recorrupted-to-Recorrupted: An Unsupervised Image Denoising Network for Arbitrary Noise Distribution

With the great breakthrough of supervised learning in the field of denoising, more and more works focus on end-to-end learning to train denoisers. In practice, however, it can be very challenging to obtain labels in support of this approach. The premise of this method is effective is that there is certain data support, but in practice, it is particularly difficult to obtain labels in the training data. Several unsupervised denoisers have emerged in recent years; however, to ensure their effectiveness, the noise model must be determined in advance, which limits the practical use of unsupervised denoising.n addition, obtaining inaccurate noise prior to noise estimation algorithms leads to low denoising accuracy. Therefore, we design a more practical denoiser that requires neither clean images as training labels nor noise model assumptions. Our method also needs the support of the noise model; the difference is that the model is generated by a residual image and a random mask during the network training process, and the input and target of the network are generated from a single noisy image and the noise model. At the same time, an unsupervised module and a pseudo supervised module are trained. The extensive experiments demonstrate the effectiveness of our framework and even surpass the accuracy of supervised denoising.

Cross-Resolution Semi-Supervised Adversarial Learning for Pansharpening

Existing deep neural network (DNN)-based methods have produced good pansharpened images. However, supervised DNN-based pansharpening methods suffer from performance degradation when fusing low spatial resolution multispectral (LR MS) and panchromatic (PAN) images at full resolution. Unsupervised DNN-based methods alleviate the issue, but their training becomes difficult owing to the absence of reference images. This paper establishes a novel semi-supervised framework to jointly learn the reconstructions of fused images from reduced- and full-resolution datasets. Specifically, we propose a cross-resolution semi-supervised adversarial learning network (CrossNet), which is composed of a supervised module and an unsupervised module. In these two modules, reduced- or full-resolution source images are disentangled as resolution-invariant components and resolution-aware components by reconstructing the fused images. Moreover, cross-resolution fused images are synthesized to enhance the disentanglement of the two kinds of components. Through the reconstruction of cross-resolution fused images, supervised and unsupervised modules are also coupled efficiently. Then, the semi-supervised framework can simultaneously make use of the supervised information in reduced-resolution datasets and mitigate the performance degradation via full-resolution datasets. Besides, adversarial learning is employed to improve the consistency between resolution-invariant components of source images at different resolutions. Finally, extensive experiments on QuickBird, GeoEye-1, WorldView-2 and WorldView-3 datasets demonstrate that the proposed CrossNet can produce state-of-the-art fusion results in terms of qualitative and quantitative evaluations. The source code is available at https://github.com/RSMagneto/CrossNet.

Semi-Supervised Domain-Adaptive Seizure Prediction via Feature Alignment and Consistency Regularization

The inter-patient variability still poses a great challenge for the real-world application of EEG-based seizure prediction, where most previous methods could only work under the patient-specific fashion and fail to generalize across patients. To address this issue, some latest studies applied supervised domain adaptation (SDA), utilizing limited labeled data from the target patient to calibrate the model. However, as the epileptic EEG representation even varies within one single patient, limited target data could hardly cover the data distribution, thus the model still shows poor generalization on the target patient. To this end, we introduce a novel semi-supervised domain adaptive seizure prediction model (SSDA-SPM), using limited labeled target data and extra unlabeled target data for adaptation. SSDA-SPM mainly consists of two unsupervised modules, namely feature alignment (FA) module and consistency regularization (CR) module. The FA module aims to transfer knowledge from existing patients to the target patient coarsely by globally aligning the data distribution between them. Then the CR module further enhances the discriminability on the target patient by pushing the decision boundary into the low-density area. Our proposed method achieves 88.8% sensitivity, 0.182/h false prediction rate (FPR) and 0.849 AUC on the CHB-MIT database and 75.7% sensitivity, 0.165/h FPR and 0.763 AUC on the Kaggle database. Experimental results demonstrate that our method has provided a promising solution to improve the cross-patient generalization for seizure prediction.

A Multi-Module Machine Learning Approach to Detect Tax Fraud

Tax fraud is one of the substantial issues affecting governments around the world. It is defined as the intentional alteration of information provided on a tax return to reduce someone’s tax liability. This is done by either reducing sales or increasing purchases. According to recent studies, governments lose over $500 billion annually due to tax fraud. A loss of this magnitude motivates tax authorities worldwide to implement efficient fraud detection strategies. Most of the work done in tax fraud using machine learning is centered on supervised models. A significant drawback of this approach is that it requires tax returns that have been previously audited, which constitutes a small percentage of the data. Other strategies focus on using unsupervised models that utilize the whole data when they search for patterns, though ignore whether the tax returns are fraudulent or not. Therefore, unsupervised models are limited in their usefulness if they are used independently to detect tax fraud. The work done in this paper focuses on addressing such limitations by proposing a fraud detection framework that utilizes supervised and unsupervised models to exploit the entire set of tax returns. The framework consists of four modules: A supervised module, which utilizes a tree-based model to extract knowledge from the data; an unsupervised module, which calculates anomaly scores; a behavioral module, which assigns a compliance score for each taxpayer; and a prediction module, which utilizes the output of the previous modules to output a probability of fraud for each tax return. We demonstrate the effectiveness of our framework by testing it on existent tax returns provided by the Saudi tax authority.

The IoT-based real-time image processing for animal recognition and classification using deep convolutional neural network (DCNN)

Human-Wildlife Conflict (HWC) is one of the crucial threats in Valparai and the surrounding region in Anamalai Tiger Reserves (ATR), Tamil Nadu, India. In this work, we projected an automatic unsupervised module for animal detection and discrimination from the animal image obtained from the camera. The captured animal images are stored in .jpeg format and are sent to the base station via Radio Frequency (RF) network. We have implemented the Deep Convolutional Neural Network (DCNN) for performing the task of feature extraction and classification. Once an animal is identified, this system will send an alert message through the Global System for Mobile Communication (GSM) to the mobile numbers of Ranger, Local people, and FM/Radio channels stored in the system. Here, the Internet of Things (IoT) plays a significant role where the results will be fed into a web page so forest officials can use this information whenever needed. This system will help to save people and wildlife.

Processing and classification of landsat and sentinel images for oil palm plantation detection

The increasing demand for remote sensing, along with the advancement of technology, has led to the development of robust, sensible, and user-friendly products that can utilise remotely captured images. Remote sensing in agriculture has gained a lot of interest recently, especially in plantation management. This technology is useful for controlling and monitoring various aspects of the plantations. One of the capabilities of remote sensing is the detection of oil palm plantations. Therefore, this paper attempts to determine the best methods for image classification, especially for land cover classification of oil palm plantations. It first focuses on the correction algorithm needed to estimate the true surface reflectance value of the satellite image data before the image is filtered to reduce any noise. The process includes the analysis of both supervised and unsupervised modules in terms of their contrast visual and reflectance spectral curve to find the best method of extracting the images’ features. In distinguishing oil palm trees, optimisation of the pre-processing of the images enables the extraction of useful information based on its spectral signature, before they are utilised as an input for the soft computing method. The results show that Artificial Neural Network (ANN) performed the best image classification with the highest overall accuracy and kappa coefficient compared to other supervised classifications. The parameters for ANN were later adjusted to identify the best ANN classification, resulting in an overall accuracy of 98.2857% and 0.9792 of kappa coefficient, and manages to effectively detect oil palm trees from the background.

In search of a suitable method for disambiguation of word senses in Bengali

The paper presents a study on word sense disambiguation (WSD) in Bengali, one of the less resourced Indian languages. The overall work is carried out in two sequential phases. In the first phase, four well-known approaches, which are often applied for sense disambiguation of words, are studied using the traditional methods. In the course of application, suitable modifications are made as well as implemented for eliciting desired results. In the second stage, a combined approach is proposed based on the results obtained from initial experiments. Within ‘supervised module’ the four commonly used methods, namely, the Decision Tree (DT) method, Support Vector Machine (SVM) method, Artificial Neural Network (ANN) method, and the Naive Bayes (NB) method are used at the baseline for the purpose of classification of senses. These baseline strategies produced 63.84%, 76.9%, 76.23%, and 80.23% accurate results, respectively, when these methods are tested on 13 mostly used Bengali ambiguous words retrieved from a Bengali text corpus. Next, two major modifications are applied on these baseline strategies to increase the level of accuracy: (a) incorporation of Lemmatization process in the system (that produces 68.30%, 79%, 78.23%, and 82.30% accurate results, respectively), and (b) operation of Bootstrapping on the systems (including lemmatization feature), which produces 70.92%, 79.15%, 79.53%, and 83% accuracy, respectively. Next, in a knowledge-based method, the traditional Lesk algorithm is implemented at the baseline which produces 31% accurate result in sense disambiguation. This strategy is further modified by Context Expansion (CE) method in the sentences using the Bengali WordNet to produce 75% accuracy. Within ‘unsupervised module’, the baseline strategy produced a 36.2% accurate result in sense disambiguation task. To enhance the level of performance, two modifications are adopted over this baseline strategy: (a) Principal Component Analysis (PCA) over the feature vector, which produces 51.2% accuracy in WSD task, and (b) Context Expansion of the sentences using the Bengali WordNet with PCA, which produces 61% accuracy in sense disambiguation task. Finally, a combined approach is adopted after considering all the effective aspects of the three methods, and it produces the highest level accuracy (92%) in the task of sense disambiguation.

Visual crowding and category specific deficits for pictorial stimuli: A neural network model

This paper describes a series of modular neural network simulations of visual object processing. In a departure from much previous work in this domain, the model described here comprises both supervised and unsupervised modules and processes real pictorial representations of items from different object categories. The unsupervised module carries out bottom-up encoding of visual stimuli, thereby developing a “perceptual” representation of each presented picture. The supervised component then classifies each perceptual representation according to a target semantic category. Model performance was assessed (1) during learning, (2) under generalisation to novel instances, and (3) after lesion damage at different stages of processing. Strong category effects were observed throughout the different experiments, with living things and musical instruments eliciting greater recognition failures relative to other categories. This pattern derives from within-category similarity effects at the level of perceptual representation and our data support the view that visual crowding can be a potentially important factor in the emergence of some category-specific impairments. The data also accord with the cascade model of object recognition, since increased competition between perceptual representations resulted in category-specific impairments even when the locus of damage was within the semantic component of the model. Some strengths and limitations of this modelling approach are discussed and the results are evaluated against some other accounts of category-specific recognition failure.

Visual crowding and category specific deficits for pictorial stimuli: A neural network model

This paper describes a series of modular neural network simulations of visual object processing. In a departure from much previous work in this domain, the model described here comprises both supervised and unsupervised modules and processes real pictorial representations of items from different object categories. The unsupervised module carries out bottom-up encoding of visual stimuli, thereby developing a “perceptual” representation of each presented picture. The supervised component then classifies each perceptual representation according to a target semantic category. Model performance was assessed (1) during learning, (2) under generalisation to novel instances, and (3) after lesion damage at different stages of processing. Strong category effects were observed throughout the different experiments, with living things and musical instruments eliciting greater recognition failures relative to other categories. This pattern derives from within-category similarity effects at the level of perceptual representation and our data support the view that visual crowding can be a potentially important factor in the emergence of some category-specific impairments. The data also accord with the cascade model of object recognition, since increased competition between perceptual representations resulted in category-specific impairments even when the locus of damage was within the semantic component of the model. Some strengths and limitations of this modelling approach are discussed and the results are evaluated against some other accounts of category-specific recognition failure.