Multi-stage Learning Research Articles

A Novel Adaptive Gain Strategy for Stochastic Learning Control

This article studies the conflicting goals of high-precision tracking and quick convergence speed, which is a longstanding problem in the learning control of stochastic systems. In such systems, a decreasing gain sequence is necessary to ensure the asymptotic convergence of the generated input sequence to a fixed limit. However, the convergence speed is adversely affected by gain sequences of this nature. In this article, we propose a novel multistage learning control strategy to resolve this conflict, where each stage consists of several iterations. The learning gain remains constant in each stage but is reduced at the transition from a given stage to the subsequent stage. The switching iteration between two stages is determined by the tracking performance index of the contracted input error and the accumulated noise drift. Furthermore, an improved mechanism is proposed to optimize the lengths of the different stages. The asymptotic convergence of the input sequence generated by the newly proposed strategy is strictly established by thoroughly analyzing the properties of the proposed gain sequence. Numerical simulations are presented to verify the theoretical results.

Hybrid Robotic Grasping With a Soft Multimodal Gripper and a Deep Multistage Learning Scheme

Grasping has long been considered an important and practical task in robotic manipulation. Yet achieving robust and efficient grasps of diverse objects is challenging, since it involves gripper design, perception, control and learning, etc. Recent learning-based approaches have shown excellent performance in grasping a variety of novel objects. However, these methods either are typically limited to one single grasping mode, or else more end effectors are needed to grasp various objects. In addition, gripper design and learning methods are commonly developed separately, which may not adequately explore the ability of a multimodal gripper. In this paper, we present a deep reinforcement learning (DRL) framework to achieve multistage hybrid robotic grasping with a new soft multimodal gripper. A soft gripper with three grasping modes (i.e., enveloping, sucking, and enveloping_then_sucking) can both deal with objects of different shapes and grasp more than one object simultaneously. We propose a novel hybrid grasping method integrated with the multimodal gripper to optimize the number of grasping actions. We evaluate the DRL framework under different scenarios (i.e., with different ratios of objects of two grasp types). The proposed algorithm is shown to reduce the number of grasping actions (i.e., enlarge the grasping efficiency, with maximum values of 161% in simulations and 154% in real-world experiments) compared to single grasping modes.

A hybrid agent-based model integrated with a multi-stage learning-based fuzzy cognitive map

PurposeThis article proposes a novel hybrid simulation model for understanding the complex tobacco use behavior.Design/methodology/approachThe model is developed by embedding the concept of the multistage learning-based fuzzy cognitive map (FCM) into the agent-based model (ABM) in order to benefit from advantageous of each methodology. The ABM is used to represent individual level behaviors while the FCM is used as a decision support mechanism for individuals. In this study, socio-demographic characteristics of individuals, tobacco control policies, and social network effect are taken into account to reflect the current tobacco use system of Turkey. The effects of plain package and COVID-19 on tobacco use behaviors of individuals are also searched under different scenarios.FindingsThe findings indicate that the proposed model provides promising results for representing the mental models of agents. Besides, the scenario analyses help to observe the possible reactions of people to new conditions according to characteristics.Originality/valueThe proposed method combined ABM and FCM with a multi-stage learning phases for modeling a complex and dynamic social problem as close as real life. It is expected to contribute for both ABM and tobacco use literature.

Multi-target Knowledge Distillation via Student Self-reflection

Knowledge distillation is a simple yet effective technique for deep model compression, which aims to transfer the knowledge learned by a large teacher model to a small student model. To mimic how the teacher teaches the student, existing knowledge distillation methods mainly adapt an unidirectional knowledge transfer, where the knowledge extracted from different intermedicate layers of the teacher model is used to guide the student model. However, it turns out that the students can learn more effectively through multi-stage learning with a self-reflection in the real-world education scenario, which is nevertheless ignored by current knowledge distillation methods. Inspired by this, we devise a new knowledge distillation framework entitled multi-target knowledge distillation via student self-reflection or MTKD-SSR, which can not only enhance the teacher’s ability in unfolding the knowledge to be distilled, but also improve the student’s capacity of digesting the knowledge. Specifically, the proposed framework consists of three target knowledge distillation mechanisms: a stage-wise channel distillation (SCD), a stage-wise response distillation (SRD), and a cross-stage review distillation (CRD), where SCD and SRD transfer feature-based knowledge (i.e., channel features) and response-based knowledge (i.e., logits) at different stages, respectively; and CRD encourages the student model to conduct self-reflective learning after each stage by a self-distillation of the response-based knowledge. Experimental results on five popular visual recognition datasets, CIFAR-100, Market-1501, CUB200-2011, ImageNet, and Pascal VOC, demonstrate that the proposed framework significantly outperforms recent state-of-the-art knowledge distillation methods.

Making a Bird AI Expert Work for You and Me.

As powerful as fine-grained visual classification (FGVC) is, responding your query with a bird name of "Whip-poor-will" or "Mallard" probably does not make much sense. This however commonly accepted in the literature, underlines a fundamental question interfacing AI and human - what constitutes transferable knowledge for human to learn from AI? This paper sets out to answer this very question using FGVC as a test bed. Specifically, we envisage a scenario where a trained FGVC model (the AI expert) functions as a knowledge provider in enabling average people (you and me) to become better domain experts ourselves. Assuming an AI expert trained using expert human labels, we anchor our focus on asking and providing solutions for two questions: (i) what is the best transferable knowledge we can extract from AI, and (ii) what is the most practical means to measure the gains in expertise given that knowledge? We propose to represent knowledge as highly discriminative visual regions that are expert-exclusive and instantiate it via a novel multi-stage learning framework. A human study of 15,000 trials shows our method is able to consistently improve people of divergent bird expertise to recognise once unrecognisable birds. We further propose a crude but benchmarkable metric TEMI and therefore allow future efforts in this direction to be comparable to ours without the need of large-scale human studies.

Progressive fusion learning: A multimodal joint segmentation framework for building extraction from optical and SAR images

Automatic and high-precision extraction of buildings from remote sensing images has a wide range of application and importance. Optical and synthetic aperture radar (SAR) images are typical types of multimodal remote sensing data with different imaging methods. To bridge the huge gap between them and achieve high-precision joint semantic segmentation, this study proposes a progressive fusion learning framework. The framework explicitly extracts the shared features (that is, modal invariants) of multimodal images as the information medium and realizes information fusion through multistage learning. Based on this framework, we design a network called the multistage multimodal fusion network (MMFNet), which uses phase as a modal invariant to joint optical and SAR images to achieve high-precision building extraction. We conducted experiments with the Multi-Sensor All-Weather Mapping aerial dataset and the WHU-OPT-SAR_WuHan satellite dataset. This study shows MMFNet has a significant extraction effect and yields more optimized extraction of the edge details of buildings, which is improved by 0.2% to 9.5% compared to other multimodal joint segmentation methods.

High-Speed Three-Dimensional Aerial Vehicle Evasion Based on a Multi-Stage Dueling Deep Q-Network

This paper proposes a multi-stage dueling deep Q-network (MS-DDQN) algorithm to address the high-speed aerial vehicle evasion problem. High-speed aerial vehicle pursuit and evasion are an ongoing game attracting significant research attention in the field of autonomous aerial vehicle decision making. However, traditional maneuvering methods are usually not applicable in high-speed scenarios. Independent of the aerial vehicle model, the implemented MS-DDQN-based method searches for an approximate optimal maneuvering policy by iteratively interacting with the environment. Furthermore, the multi-stage learning mechanism was introduced to improve the training data quality. Simulation experiments were conducted to compare the proposed method with several typical evasion maneuvering policies and to reveal the effectiveness and robustness of the proposed MS-DDQN algorithm.

End-to-End Learning to Grasp via Sampling From Object Point Clouds

The ability to grasp objects is an essential skill that enables many robotic manipulation tasks. Recent works have studied point cloud-based methods for object grasping by starting from simulated datasets and have shown promising performance in real-world scenarios. Nevertheless, many of them still rely on ad-hoc geometric heuristics to generate grasp candidates, which fail to generalize to objects with significantly different shapes with respect to those observed during training. Several approaches exploit complex multi-stage learning strategies and local neighborhood feature extraction while ignoring semantic global information. Furthermore, they are inefficient in terms of number of training samples and time required for inference. In this letter, we propose an end-to-end learning solution to generate 6-DOF parallel-jaw grasps starting from the 3D partial view of the object. Our Learning to Grasp (L2G) method gathers information from the input point cloud through a new procedure that combines a differentiable sampling strategy to identify the visible contact points, with a feature encoder that leverages local and global cues. Overall, L2G is guided by a multi-task objective that generates a diverse set of grasps by optimizing contact point sampling, grasp regression, and grasp classification. With a thorough experimental analysis, we show the effectiveness of L2G as well as its robustness and generalization abilities.

Optimistic Temporal Difference Learning for 2048

Temporal difference (TD) learning and its variants, such as multistage TD (MS-TD) learning and temporal coherence (TC) learning, have been successfully applied to 2048. These methods rely on the stochasticity of the environment of 2048 for exploration. In this paper, we propose to employ optimistic initialization (OI) to encourage exploration for 2048, and empirically show that the learning quality is significantly improved. This approach optimistically initializes the feature weights to very large values. Since weights tend to be reduced once the states are visited, agents tend to explore those states which are unvisited or visited few times. Our experiments show that both TD and TC learning with OI significantly improve the performance. As a result, the network size required to achieve the same performance is significantly reduced. With additional tunings such as expectimax search, multistage learning, and tile-downgrading technique, our design achieves the state-of-the-art performance, namely an average score of 625 377 and a rate of 72% reaching 32768 tiles. In addition, for sufficiently large tests, 65536 tiles are reached at a rate of 0.02%.

A Reinforcement Learning Approach to Speech Coding

Speech coding is an essential technology for digital cellular communications, voice over IP, and video conferencing systems. For more than 25 years, the main approach to speech coding for these applications has been block-based analysis-by-synthesis linear predictive coding. An alternative approach that has been less successful is sample-by-sample tree coding of speech. We reformulate this latter approach as a multistage reinforcement learning problem with L step lookahead that incorporates exploration and exploitation to adapt model parameters and to control the speech analysis/synthesis process on a sample-by-sample basis. The minimization of the spectrally shaped reconstruction error to finite depth manages complexity and serves as an effective stand in for the overall subjective evaluation of reconstructed speech quality and intelligibility. Different control policies that attempt to persistently excite the system states and that encourage exploration are studied and evaluated. The resulting methods produce reconstructed speech quality competitive with the most popular speech codec utilized today. This new reinforcement learning formulation provides new insights and opens up new directions for system design and performance improvement.

Abstract 5052: Patchless deep transfer learning for improved mammographic breast mass classification

Abstract Purpose: Despite mammographic breast mass change being the most important finding in characterizing breast cancer in young women with dense breasts, a mammogram (Mg) is susceptible to false positives and false negatives in distinguishing between benign and malignant breast mass images. Additional tests such as ultrasound and biopsy might be needed to decide results. Setting aside the improvements in classifying breast mass Mg images via deep learning (DL), obtaining large training data and ensuring generalizations across different datasets with robust and well-optimized algorithms is a challenge. ImageNet based transfer learning has been utilized to address the unavailability of large datasets and robust algorithms. However, it is yet to achieve the desired accuracy, sensitivity, and specificity for DL models to be used as a standalone tool. Furthermore, previous works are computationally infeasible where exhaustive patch separation is carried out to segment the region of interest before training, which makes processing computationally complex and time-consuming. Here we propose a novel deep learning method based on multistage transfer learning from ImageNet and cancer cell line images pre-trained EfficientNetB2 model to classify mammographic breast mass as either benign or malignant. Methods: We trained our model on three publicly available datasets, 13,128 Digital Database for Screening Mammography (DDSM), 7632 INbreast, and 3816 Mammographic Image Analysis Society (MIAS) Mg breast mass images. Additionally, we trained our model on a mixed dataset of images from the three datasets to evaluate robustness. Data were sorted into 6:2:2 ratio for training, validation, and test, respectively. The microscopic cancer cell line dataset size was 38, 080 images. Results: We obtained an average 5-fold cross-validation AUC of 0.9999, test accuracy of 99.99%, sensitivity of 1, and specificity of 0.9998 on DDSM, AUC of 0.9997, test accuracy of 99.99%, sensitivity of 0.9972, and specificity of 0.9988 on INbreast, and AUC of 0.9987, test accuracy of 99.89%, sensitivity of 0.9987, specificity of 1 on MIAS, and AUC of 0.9997, test accuracy of 99.91%, sensitivity of 0.9993, and specificity of 0.9989 on the mixed dataset. Moreover, we acquired a P-value of 0.019 in the investigation of a statistically significant improvement in test accuracy from using our method compared to the conventional ImageNet based transfer learning on DDSM dataset. Conclusion: Our study suggests that utilizing cancer cell line images further improved the learning process alleviating the need for large Mg training data. Moreover, our method achieved better performance without applying the computationally complex patch separation task. The findings of this study are of crucial importance in the early diagnosis of breast cancer in young women with dense breasts where mammography struggles. Citation Format: Gelan Ayana, Jinhyung Park, Se-woon Choe. Patchless deep transfer learning for improved mammographic breast mass classification [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5052.

Fusing Multimodal and Anatomical Volumes of Interest Features Using Convolutional Auto-Encoder and Convolutional Neural Networks for Alzheimer's Disease Diagnosis.

Alzheimer’s disease (AD) is an age-related disease that affects a large proportion of the elderly. Currently, the neuroimaging techniques [e.g., magnetic resonance imaging (MRI) and positron emission tomography (PET)] are promising modalities for AD diagnosis. Since not all brain regions are affected by AD, a common technique is to study some region-of-interests (ROIs) that are believed to be closely related to AD. Conventional methods used ROIs, identified by the handcrafted features through Automated Anatomical Labeling (AAL) atlas rather than utilizing the original images which may induce missing informative features. In addition, they learned their framework based on the discriminative patches instead of full images for AD diagnosis in multistage learning scheme. In this paper, we integrate the original image features from MRI and PET with their ROIs features in one learning process. Furthermore, we use the ROIs features for forcing the network to focus on the regions that is highly related to AD and hence, the performance of the AD diagnosis can be improved. Specifically, we first obtain the ROIs features from the AAL, then we register every ROI with its corresponding region of the original image to get a synthetic image for each modality of every subject. Then, we employ the convolutional auto-encoder network for learning the synthetic image features and the convolutional neural network (CNN) for learning the original image features. Meanwhile, we concatenate the features from both networks after each convolution layer. Finally, the highly learned features from the MRI and PET are concatenated for brain disease classification. Experiments are carried out on the ADNI datasets including ADNI-1 and ADNI-2 to evaluate our method performance. Our method demonstrates a higher performance in brain disease classification than the recent studies.

Tuberculosis Bacteria Detection and Counting in Fluorescence Microscopy Images Using a Multi-Stage Deep Learning Pipeline

The manual observation of sputum smears by fluorescence microscopy for the diagnosis and treatment monitoring of patients with tuberculosis (TB) is a laborious and subjective task. In this work, we introduce an automatic pipeline which employs a novel deep learning-based approach to rapidly detect Mycobacterium tuberculosis (Mtb) organisms in sputum samples and thus quantify the burden of the disease. Fluorescence microscopy images are used as input in a series of networks, which ultimately produces a final count of present bacteria more quickly and consistently than manual analysis by healthcare workers. The pipeline consists of four stages: annotation by cycle-consistent generative adversarial networks (GANs), extraction of salient image patches, classification of the extracted patches, and finally, regression to yield the final bacteria count. We empirically evaluate the individual stages of the pipeline as well as perform a unified evaluation on previously unseen data that were given ground-truth labels by an experienced microscopist. We show that with no human intervention, the pipeline can provide the bacterial count for a sample of images with an error of less than 5%.

Selecting Post-Processing Schemes for Accurate Detection of Small Objects in Low-Resolution Wide-Area Aerial Imagery

In low-resolution wide-area aerial imagery, object detection algorithms are categorized as feature extraction and machine learning approaches, where the former often requires a post-processing scheme to reduce false detections and the latter demands multi-stage learning followed by post-processing. In this paper, we present an approach on how to select post-processing schemes for aerial object detection. We evaluated combinations of each of ten vehicle detection algorithms with any of seven post-processing schemes, where the best three schemes for each algorithm were determined using average F-score metric. The performance improvement is quantified using basic information retrieval metrics as well as the classification of events, activities and relationships (CLEAR) metrics. We also implemented a two-stage learning algorithm using a hundred-layer densely connected convolutional neural network for small object detection and evaluated its degree of improvement when combined with the various post-processing schemes. The highest average F-scores after post-processing are 0.902, 0.704 and 0.891 for the Tucson, Phoenix and online VEDAI datasets, respectively. The combined results prove that our enhanced three-stage post-processing scheme achieves a mean average precision (mAP) of 63.9% for feature extraction methods and 82.8% for the machine learning approach.

Deep Transfer Learning for Biology Cross-Domain Image Classification

Automatic biology image classification is essential for biodiversity conservation and ecological study. Recently, due to the record-shattering performance, deep convolutional neural networks (DCNNs) have been used more often in biology image classification. However, training DCNNs requires a large amount of labeled data, which may be difficult to collect for some organisms. This study was carried out to exploit cross-domain transfer learning for DCNNs with limited data. According to the literature, previous studies mainly focus on transferring from ImageNet to a specific domain or transferring between two closely related domains. While this study explores deep transfer learning between species from different domains and analyzes the situation when there is a huge difference between the source domain and the target domain. Inspired by the analysis of previous studies, the effect of biology cross-domain image classification in transfer learning is proposed. In this work, the multiple transfer learning scheme is designed to exploit deep transfer learning on several biology image datasets from different domains. There may be a huge difference between the source domain and the target domain, causing poor performance on transfer learning. To address this problem, multistage transfer learning is proposed by introducing an intermediate domain. The experimental results show the effectiveness of cross-domain transfer learning and the importance of data amount and validate the potential of multistage transfer learning.

Exploratory study on classification of chronic obstructive pulmonary disease combining multi-stage feature fusion and machine learning

BackgroundDue to the complexity and high heterogeneity of the acute exacerbation of chronic obstructive pulmonary disease (AECOPD), the guidelines (global initiative for chronic obstructive, GOLD) is unable to fully guide the treatment of AECOPD.ObjectivesTo provide a rapid treatment in line with the development of the AECOPD after admission. In this paper, we propose a multi-stage feature fusion (MSFF) framework combining machine learning to track the diseases deterioration risk of the AECOPD.MethodsFirst, we identify 408 AECOPD patients as the study population. Then, feature segment and fusion methods are applied to generate the phased data set. Finally, human studies are designed to evaluate the performance of the MSFF framework.ResultsThe experimental results show that the proposed framework is potential to obtain the full-process tracking of deterioration risk for the AECOPD patients. The proposed MSFF framework achieves a higher overall accuracy average and F1 scores than the four physician groups i.e., IM, Surgery, Emergency, and ICU.ConclusionsThe proposed MSFF model may serve as a useful disease tracking tool to estimate the deterioration risk at each stage, and finally achieve the disease monitoring and management for AECOPD patients.

A two-stage complex network using cycle-consistent generative adversarial networks for speech enhancement

Cycle-consistent generative adversarial networks (CycleGAN) have shown their promising performance for speech enhancement (SE), while one intractable shortcoming of these CycleGAN-based SE systems is that the noise components propagate throughout the cycle and cannot be completely eliminated. Additionally, conventional CycleGAN-based SE systems only estimate the spectral magnitude, while the phase is unaltered. Motivated by the multi-stage learning concept, we propose a novel two-stage denoising system that combines a CycleGAN-based magnitude enhancing network and a subsequent complex spectral refining network in this paper. Specifically, in the first stage, a CycleGAN-based model is responsible for only estimating magnitude, which is subsequently coupled with the original noisy phase to obtain a coarsely enhanced complex spectrum. After that, the second stage is applied to further suppress the residual noise components and estimate the clean phase by a complex spectral mapping network, which is a pure complex-valued network composed of complex 2D convolution/deconvolution and complex temporal-frequency attention blocks. Experimental results on two public datasets demonstrate that the proposed approach consistently surpasses previous one-stage CycleGANs and other state-of-the-art SE systems in terms of various evaluation metrics, especially in background noise suppression.

A hybrid multi-stage learning technique based on brain storming optimization algorithm for breast cancer recurrence prediction

Breast cancer disease is considered to be the second leading reason for death among women. Unfortunately, even if the treatment of cancer started soon after diagnosis, the cancer cells may remain in the body, and cancer may recur. Various Machine Learning (ML) methods to predict breast cancer recurrence were applied recently, and the ML methods’ performance needs to be examined to determine the proper method for prediction. Usually, the datasets contain many features which may sometimes mislead the prediction process; as some features may lead to confusion or inaccurate prediction. Thus, in this study, two breast cancer recurrence datasets were statistically analyzed and further refined by Brain Storming Optimization algorithm (BSO). The proposed multi-stages technique consists of three main stages; first, the statistical feature selection methods (SFM) which statistically select the discriminative features based on importance ranking and features correlations and statistical hypothesis testing, to be passed to the second stage. The features are ranked based on their correlation with the class variable. The second stage namely the multi classifier (MC), which evaluates each method based on three classifiers and produces a combination of features performed by two SFM and three classifiers. In the third stage, the best combination of the selected features was recognized by the BSO algorithm to search for an optimal solution that produces the highest accuracy. In addition, the BSO algorithm has been modified to deal with the feature selection problem. The performance of the proposed technique was evaluated by stratified 10-fold cross-validation. As a result, the multi-stage learning technique showed to be effective in ranking the features and improved the classification accuracy for breast cancer recurrence.

Speech Enhancement Using Multi-Stage Self-Attentive Temporal Convolutional Networks

Multi-stage learning is an effective technique to invoke multiple deep-learning modules sequentially. This paper applies multi-stage learning to speech enhancement by using a multi-stage structure, where each stage comprises a self-attention (SA) block followed by stacks of temporal convolutional network (TCN) blocks with doubling dilation factors. Each stage generates a prediction that is refined in a subsequent stage. A fusion block is inserted at the input of later stages to re-inject original information. The resulting multi-stage speech enhancement system, in short, multi-stage SA-TCN, is compared with state-of-the-art deep-learning speech enhancement methods using the LibriSpeech and VCTK data sets. The multi-stage SA-TCN system's hyper-parameters are fine-tuned, and the impact of the SA block, the fusion block and the number of stages are determined. The use of a multi-stage SA-TCN system as a front-end for automatic speech recognition systems is investigated as well. It is shown that the multi-stage SA-TCN systems perform well relative to other state-of-the-art systems in terms of speech enhancement and speech recognition scores.

Construction of Longitudinally Consistent 4D Infant Cerebellum Atlases Based on Deep Learning.

Longitudinal infant dedicated cerebellum atlases play a fundamental role in characterizing and understanding the dynamic cerebellum development during infancy. However, due to the limited spatial resolution, low tissue contrast, tiny folding structures, and rapid growth of the cerebellum during this stage, it is challenging to build such atlases while preserving clear folding details. Furthermore, the existing atlas construction methods typically independently build discrete atlases based on samples for each age group without considering the within-subject temporal consistency, which is critical for large-scale longitudinal studies. To fill this gap, we propose an age-conditional multi-stage learning framework to construct longitudinally consistent 4D infant cerebellum atlases. Specifically, 1) A joint affine and deformable atlas construction framework is proposed to accurately build temporally continuous atlases based on the entire cohort, and rapidly warp the new images to the atlas space; 2) A longitudinal constraint is employed to enforce the within-subject temporal consistency during atlas building; 3) A Correntropy based regularization loss is further exploited to enhance the robustness of our framework. Our atlases are constructed based on 405 longitudinal scans from 187 healthy infants with age ranging from 6 to 27 months, and are compared to the atlases built by state-of-the-art algorithms. Results demonstrate that our atlases preserve more structural details and fine-grained cerebellum folding patterns, which ensure higher accuracy in subsequent atlas-based registration and segmentation tasks.