Small Dataset Size Research Articles

Transferable machine learning model for the aerodynamic prediction of swept wings

With their development, machine learning models can be used instead of computational fluid dynamics simulations to predict flow fields in aerodynamic optimization. However, it is difficult to construct a prediction model for swept wings with various planform geometries because too many samples are required to cover the parameter space. In the present paper, a new model framework is proposed to predict wing surface pressure and friction distributions with fewer samples. The distributed geometry parameters along spanwise are used as model inputs instead of the global planform parameters, and processors are designed to help the model better learn the local effect of geometric variation. The model is trained and tested on simple swept wings with single segment and linear twist distribution, where it outperforms the global input model by 57.6% in terms of lift coefficient prediction errors on small dataset sizes. The distributed input also enables the model to be transferred from single wings to more engineering-practical yet complex kink wings. After fine-tuning with a few samples, model accuracy for kink wings can be similar to that of simple wings, which proves the model for wings with complex planform geometries can be efficiently built with the proposed method.

Patch-and-amplify Capsule Network for the recognition of gastrointestinal diseases

Deep learning (DL) algorithms require massive amounts of data in diverse variations to attain close to human recognition accuracy. This enables them to perform well on unseen in-distribution viewpoints. However, critical areas such as health face a challenge in generating large data. This is mainly attributed to regulatory policies surrounding the collection of such data. Hence, the domain of health is known to have smaller datasets. This limitation, therefore, requires algorithms that can derive maximum benefit from the small-size dataset. An example of such an algorithm is Capsule Network (CapsNet). Though Capsule Networks perform well on smaller datasets, they attain low performance on datasets consisting of complex real-life images. This challenge is attributed to their weak encoders. This paper, therefore, proposes a feature enhancement technique termed patch-and-amplify to improve the feature extraction task and enhance extracted features in the encoder layer. A model-specific squash function is proposed to enhance the coupling of capsules between the higher and lower capsule layers and reduce the impact of polarization. To evaluate the network's robustness and generalizability, the proposed model was evaluated on CIFAR-10, Fashion-MNIST, and Kvasir-V2 datasets achieving recognition accuracies of 85.50 %, 98.40 %, and 93.40 %. Extensive visualizations were explored to assist in understanding the predictions and internal structure of the network. The proposed model performs well compared with the state-of-the-art in the literature and can serve as a helpful tool in assisting medical practitioners in diagnosing gastrointestinal diseases.

Harmonizing heterogeneous transcriptomics datasets for machine learning-based analysis to identify spaceflown murine liver-specific changes

NASA has employed high-throughput molecular assays to identify sub-cellular changes impacting human physiology during spaceflight. Machine learning (ML) methods hold the promise to improve our ability to identify important signals within highly dimensional molecular data. However, the inherent limitation of study subject numbers within a spaceflight mission minimizes the utility of ML approaches. To overcome the sample power limitations, data from multiple spaceflight missions must be aggregated while appropriately addressing intra- and inter-study variabilities. Here we describe an approach to log transform, scale and normalize data from six heterogeneous, mouse liver-derived transcriptomics datasets (ntotal = 137) which enabled ML-methods to classify spaceflown vs. ground control animals (AUC ≥ 0.87) while mitigating the variability from mission-of-origin. Concordance was found between liver-specific biological processes identified from harmonized ML-based analysis and study-by-study classical omics analysis. This work demonstrates the feasibility of applying ML methods on integrated, heterogeneous datasets of small sample size.

DPML: Prior-guided multitask learning for dental object recognition on limited panoramic radiograph dataset

Addressing the growing need for advanced computer-aided diagnostic tools in dentistry, this study presents DPML (Dental Prior-guided Multitask Learning), a pioneering framework tailored for the analysis of dental panoramic radiographs in the face of data scarcity. DPML introduces a novel multitask learning strategy that effectively mitigates the limitations of sparse, well-annotated dental datasets. The methodology integrates four advanced techniques: targeted data augmentation for enhanced image diversity, segmentation structures for improved edge definition, polar-coordinate methods for precise tooth contour extraction, and a spatially informed classification strategy for accurate tooth type identification. Our comparative analysis demonstrates that DPML markedly surpasses traditional approaches in tooth segmentation, detection, and classification. Remarkably, it achieves a 95.06% mean Average Precision (mAP) and a 97.70% precision rate in dental object recognition tasks on a small-size dataset.

Uncertainty estimation using a 3D probabilistic U-Net for segmentation with small radiotherapy clinical trial datasets

Background and objectivesBio-medical image segmentation models typically attempt to predict one segmentation that resembles a ground-truth structure as closely as possible. However, as medical images are not perfect representations of anatomy, obtaining this ground truth is not possible. A surrogate commonly used is to have multiple expert observers define the same structure for a dataset. When multiple observers define the same structure on the same image there can be significant differences depending on the structure, image quality/modality and the region being defined. It is often desirable to estimate this type of aleatoric uncertainty in a segmentation model to help understand the region in which the true structure is likely to be positioned. Furthermore, obtaining these datasets is resource intensive so training such models using limited data may be required. With a small dataset size, differing patient anatomy is likely not well represented causing epistemic uncertainty which should also be estimated so it can be determined for which cases the model is effective or not. MethodsWe use a 3D probabilistic U-Net to train a model from which several segmentations can be sampled to estimate the range of uncertainty seen between multiple observers. To ensure that regions where observers disagree most are emphasised in model training, we expand the Generalised Evidence Lower Bound (ELBO) with a Constrained Optimisation (GECO) loss function with an additional contour loss term to give attention to this region. Ensemble and Monte-Carlo dropout (MCDO) uncertainty quantification methods are used during inference to estimate model confidence on an unseen case. We apply our methodology to two radiotherapy clinical trial datasets, a gastric cancer trial (TOPGEAR, TROG 08.08) and a post-prostatectomy prostate cancer trial (RAVES, TROG 08.03). Each dataset contains only 10 cases each for model development to segment the clinical target volume (CTV) which was defined by multiple observers on each case. An additional 50 cases are available as a hold-out dataset for each trial which had only one observer define the CTV structure on each case. Up to 50 samples were generated using the probabilistic model for each case in the hold-out dataset. To assess performance, each manually defined structure was matched to the closest matching sampled segmentation based on commonly used metrics. ResultsThe TOPGEAR CTV model achieved a Dice Similarity Coefficient (DSC) and Surface DSC (sDSC) of 0.7 and 0.43 respectively with the RAVES model achieving 0.75 and 0.71 respectively. Segmentation quality across cases in the hold-out datasets was variable however both the ensemble and MCDO uncertainty estimation approaches were able to accurately estimate model confidence with a p-value < 0.001 for both TOPGEAR and RAVES when comparing the DSC using the Pearson correlation coefficient. ConclusionsWe demonstrated that training auto-segmentation models which can estimate aleatoric and epistemic uncertainty using limited datasets is possible. Having the model estimate prediction confidence is important to understand for which unseen cases a model is likely to be useful.

An Effective Classification of Brain Tumor using Deep Learning Techniques

Abstract: Brain tumor classification is a critical task in medical imaging, aiding in timely diagnosis and treatment planning. In this paper, we propose a comprehensive approach utilizing deep learning models for the classification of brain tumors into four categories: glioma, meningioma, pituitary tumors, and no tumors. We employ state-of-the-art convolutional neural network (CNN) architectures including ResNet50, DenseNet201, EfficientNetB3, and InceptionV3 for the classification task. To enhance the performance of the model, we employ resizing and augmentation techniques such as flips and rotation, thereby increasing the diversity of the training dataset. This is particularly crucial due to the limitations posed by small-sized datasets in previous methodologies. Our findings underscore the efficacy of deep learning approaches in brain tumor classification, with EfficientNetB3 emerging as a promising model for accurate diagnosis. Furthermore, our utilization of resizing and augmentation techniques demonstrates their significance in mitigating the challenges associated with limited training data.

An effective ensemble learning approach for classification of glioma grades based on novel MRI features

The preoperative diagnosis of brain tumors is important for therapeutic planning as it contributes to the tumors’ prognosis. In the last few years, the development in the field of artificial intelligence and machine learning has contributed greatly to the medical area, especially the diagnosis of the grades of brain tumors through radiological images and magnetic resonance images. Due to the complexity of tumor descriptors in medical images, assessing the accurate grade of glioma is a major challenge for physicians. We have proposed a new classification system for glioma grading by integrating novel MRI features with an ensemble learning method, called Ensemble Learning based on Adaptive Power Mean Combiner (EL-APMC). We evaluate and compare the performance of the EL-APMC algorithm with twenty-one classifier models that represent state-of-the-art machine learning algorithms. Results show that the EL-APMC algorithm achieved the best performance in terms of classification accuracy (88.73%) and F1-score (93.12%) over the MRI Brain Tumor dataset called BRATS2015. In addition, we showed that the differences in classification results among twenty-two classifier models have statistical significance. We believe that the EL-APMC algorithm is an effective method for the classification in case of small-size datasets, which are common cases in medical fields. The proposed method provides an effective system for the classification of glioma with high reliability and accurate clinical findings.

Strategic Selection of Machine Learning Models for Short-term Trading Optimization

Stock market prediction is a key aspect for a financial analyst and investor, as it facilitates better decisions and improved investment strategies. Towards this end, this project compares how machine learning (ML) models perform in predicting stock market prices with relatively small dataset sizes so as to get the best performing one. This study evaluates the efficiency and accuracy of artificial neural networks, support vector regression, LSTM and decision trees in capturing market trends and forecasting stock movements. The importance of our proposed paper lies in its potential to demystify the complexities of financial markets for students and new entrants in the field of finance. Participants can garner practical insights on how the stock market operates and the application of theoretical ML concepts to financial analytics by identifying the best-performing ML models with limited datasets, which is a common occurrence in real-life situations. This paper aims to analyse the commonly present machine learning models in stock market prediction and choose the most effective model depending upon their performance, which is achieved using limited information (short term data). This would provide an edge in taking small (less-risky) informed financial decisions without extensive study of the stock market.

An approach to estimate the low cycle fatigue probabilistic curves of PBF-LB/M 316L steel from small size datasets using the remora optimization algorithm

The significant dispersion in fatigue properties is a prevalent attribute observed in metals manufactured by the Additive Manufacturing (AM). To ascertain the safety of components subjected to low cycle fatigue (LCF) loadings at a temperature of 550 °C, it is necessary to carry out a fatigue reliability assessment of AM metals. Nevertheless, the number of LCF life tests is limited owing to time-consuming nature of test and high total production cost of AM materials. Such a small size dataset cannot satisfy the criteria of the traditional reliability assessment method. This work aims to develop a novel method to evaluate the LCF reliability of 316L stainless steel manufactured by laser-based powder bed fusion of metals (PBF-LB/M) technology. Firstly, the Weibull model is coupled with the Coffin-Manson-Basquin relationship in order to characterize the dispersion of life data points at various strain levels. Secondly, reliability results from a small size dataset show a maximum relative error in fatigue life of less than 25 %, which is similar to results from a large size dataset and the traditional method using fatigue data from public sources. Among this step, the remora optimization algorithm is firstly applied to calculate the characteristic parameters of newly-proposed method. Thirdly, the LCF reliability of PBF-LB/M 316L based on the small size dataset is evaluated. The N95% is equivalent to 44 % of N50%. Finally, the PBF-LB/M 316L is compared with traditional 316L considering the reliability. The LCF properties of PBF-LB/M 316L is similar to the traditional 316L at 50 % reliability, however it’s slightly worse at 95 % reliability.

Machine learning for determination of activity of water and activity coefficients of electrolytes in binary solutions

Activity of water and electrolytes in aqueous solutions is of utmost importance for multiple industrial applications. However, experimental determination of such values is time-consuming, while calculation of activity coefficients using numerical methods is challenging. By training neural networks models on literature data, one could predict activity of water and electrolytes easily, without requiring any experiment. In this paper, multiple descriptors (or features) are compared to predict activity coefficients of electrolytes and activity of water in electrolyte solutions. A neural network based on the Levenberg-Marquardt algorithm (LM-NN) showed high accuracy to calculate values, despite the small size of the training datasets. Both activity coefficients of electrolytes and activity of water in electrolyte solutions can be predicted accurately even on unseen data, using simple descriptors such as electrolyte concentration, ion sizes and charges. However, some discrepancies were observed due to the lack of representativeness of the training dataset. This could be solved by selecting training data sets that are similar (e.g. same group of the periodic table) to the unknown values, or by including available experimental data for the salt considered. The ability of the LM-NN to solve non-linear least square curve fitting problems makes it a good candidate to fit experimental activity coefficient data, with the advantage of simplicity as compared to e-NRTL or UNIQUAC methods. This method paves the way for accurate and quick determination of thermodynamic data for electrolyte solutions (and beyond) using machine learning, without necessitating large training datasets.

Enhancing performance of vision transformers on small datasets through local inductive bias incorporation

Vision transformers (ViTs) achieve remarkable performance on large datasets, but tend to perform worse than convolutional neural networks (CNNs) when trained from scratch on smaller datasets, possibly due to a lack of local inductive bias in the architecture. Recent studies have therefore added locality to the architecture and demonstrated that it can help ViTs achieve performance comparable to CNNs in the small-size dataset regime. Existing methods, however, are architecture-specific or have higher computational and memory costs. Thus, we propose a module called Local InFormation Enhancer (LIFE) that extracts patch-level local information and incorporates it into the embeddings used in the self-attention block of ViTs. Our proposed module is memory and computation efficient, as well as flexible enough to process auxiliary tokens such as the classification and distillation tokens. Empirical results show that the addition of the LIFE module improves the performance of ViTs on small image classification datasets. We further demonstrate how the effect can be extended to downstream tasks, such as object detection and semantic segmentation. In addition, we introduce a new visualization method, Dense Attention Roll-Out, specifically designed for dense prediction tasks, allowing the generation of class-specific attention maps utilizing the attention maps of all tokens. The code for this project is available on Github (https://github.com/NeurAI-Lab/LIFEhttps://github.com/NeurAI-Lab/LIFE).

Sparse L0-norm least squares support vector machine with feature selection

Least squares support vector machine (LSSVM) is a powerful classification tool based on hyperplanes. But the classical LSSVM does not perform well on small sample size data sets (SSS) because it lacks feature selection capability. To address this issue, we introduce an L0-norm regularization term into the objective function of LSSVM, and propose a new sparse LSSVM model called L0-LSSVM. However, the introduction of the L0-norm makes the resulting optimization problem nonconvex and nonsmooth. To overcome these challenges, instead of using the traditional convex or non-convex approximation of the L0-norm, we design an efficient algorithm for L0-LSSVM based on alternating direction method of multipliers (ADMM) where the L0-norm is handled by employing the proximity operator. Numerical results demonstrate that L0-LSSVM not only exhibits excellent generalization performance and feature selection capability but also offers a substantial improvement in computational speed.

Use of 3D-CAPSNET and RNN models for 4D fMRI-based Alzheimer’s Disease Pre-detection

Predicting Alzheimer's disease (AD) at an early stage can assist more successfully prevent cognitive decline. Numerous investigations have focused on utilizing various convolutional neural network (CNN)-based techniques for automated diagnosis of AD through resting-state functional magnetic resonance imaging (rs-fMRI). Two main constraints face the methodologies presented in these studies. First, overfitting occurs due to the small size of fMRI datasets. Second, an effective modeling of the 4D information from fMRI sessions is required. In order to represent the 4D information, some studies used the deep learning techniques on functional connectivity matrices created from fMRI data, or on fMRI data as distinct 2D slices or 3D volumes. However, this results in information loss in both types of methods. In order to model the spatiotemporal (4D) information of fMRI data for AD diagnosis, a new model based on the capsule network (CapsNet) and recurrent neural network (RNN) is proposed in this study. To assess the suggested model's effectiveness, experiments were run. The findings show that the suggested model could classify AD against normal control (NC) and late mild cognitive impairment (lMCI) against early mild cognitive impairment (eMCI) with accuracy rates of 94.5% and 61.8%, respectively.

Sentiment analysis on a low-resource language dataset using multimodal representation learning and cross-lingual transfer learning

Affect Sensing is a rapidly growing field with the potential to revolutionize human–computer interaction, healthcare, and many more applications. Multimodal Sentiment Analysis (MSA) is a recent research area that exploits the multimodal nature of video data for affect sensing. However, the success of a multimodal framework depends on addressing the challenges associated with integrating diverse modalities and selecting informative features. We propose a novel multimodal representation learning framework using multimodal autoencoders that learns a comprehensive representation of the underlying heterogeneous modalities. Affect Sensing is even more challenging in low-resource languages because annotated video datasets and language-specific models are limited. To address this concern, we introduce Multimodal Sentiment Analysis Corpus in Tamil (MSAT), a small-sized dataset in the Tamil language for MSA, and exhibit how a novel technique involving cross-lingual transfer learning in a multimodal setting, leverages the knowledge gained by training the model on a larger English MSA dataset to fine-tune a much smaller Tamil MSA dataset. Our transfer learning model achieves significant gain in the Tamil dataset by a large margin. Our experiments demonstrate that we can build efficient, generalized models for low-resource languages by using the existing MSA datasets.

Adaptive Stacking Ensemble Techniques for Early Severity Classification of COVID-19 Patients

During outbreaks of infectious diseases, such as COVID-19, it is critical to rapidly determine treatment priorities and identify patients requiring hospitalization based on clinical severity. Although various machine learning models have been developed to predict COVID-19 severity, most have limitations, such as small dataset sizes, the limited availability of clinical variables, or a constrained classification of severity levels by a single classifier. In this paper, we propose an adaptive stacking ensemble technique that identifies various COVID-19 patient severity levels and separates them into three formats: Type 1 (low or high severity), Type 2 (mild, severe, critical), and Type 3 (asymptomatic, mild, moderate, severe, fatal). To enhance the model’s generalizability, we utilized a nationwide dataset from the South Korean government, comprising data from 5644 patients across over 100 hospitals. To address the limited availability of clinical variables, our technique employs data-driven strategies and a proposed feature selection method. This ensures the availability of clinical variables across diverse hospital environments. To construct optimal stacking ensemble models, our technique adaptively selects candidate base classifiers by analyzing the correlation between their predicted outcomes and performance. It then automatically determines the optimal multi-layer combination of base and meta-classifiers using a greedy search algorithm. To further improve the performance, we applied various techniques, including imputation of missing values and oversampling. The experimental results demonstrate that our stacking ensemble models significantly outperform existing single classifiers and AutoML approaches, with improvements of 6.42% and 8.86% in F1 and AUC scores for Type 1, 9.59% and 6.68% for Type 2, and 11.94% and 9.24% for Type 3, respectively. Consequently, our approach improves the prediction of COVID-19 severity levels and potentially assists frontline healthcare providers in making informed decisions.

Improving Transferability for Cross-Domain Trajectory Prediction via Neural Stochastic Differential Equation

Multi-agent trajectory prediction is crucial for various practical applications, spurring the construction of many large-scale trajectory datasets, including vehicles and pedestrians. However, discrepancies exist among datasets due to external factors and data acquisition strategies. External factors include geographical differences and driving styles, while data acquisition strategies include data acquisition rate, history/prediction length, and detector/tracker error. Consequently, the proficient performance of models trained on large-scale datasets has limited transferability on other small-size datasets, bounding the utilization of existing large-scale datasets. To address this limitation, we propose a method based on continuous and stochastic representations of Neural Stochastic Differential Equations (NSDE) for alleviating discrepancies due to data acquisition strategy. We utilize the benefits of continuous representation for handling arbitrary time steps and the use of stochastic representation for handling detector/tracker errors. Additionally, we propose a dataset-specific diffusion network and its training framework to handle dataset-specific detection/tracking errors. The effectiveness of our method is validated against state-of-the-art trajectory prediction models on the popular benchmark datasets: nuScenes, Argoverse, Lyft, INTERACTION, and Waymo Open Motion Dataset (WOMD). Improvement in performance gain on various source and target dataset configurations shows the generalized competence of our approach in addressing cross-dataset discrepancies.

Abstract 7380: Predicting immunotherapy outcomes from H&amp;E images in lung cancer

Abstract Checkpoint blockade immunotherapy is a cornerstone of lung cancer treatment, but there is a need to improve the identification of patients who will respond favorably. Here, we explored a deep learning approach to predict immunotherapy outcomes from hematoxylin and eosin (H&amp;E) images in non-small cell lung cancer (NSCLC). We included 150 unique cases with metastatic NSCLC (113 adenocarcinoma, 29 squamous cell, 8 other) treated with anti-PD-1/PD-L1 immunotherapy (56 nivolumab, 49 atezolizumab, 44 pembrolizumab, 1 durvalumab) as mono or combination (14 with chemotherapy, 1 with ipilimumab) therapy in a single institution. Each case consisted of a representative H&amp;E whole slide image (53 biopsies, 50 needle core biopsies, 47 resections) obtained prior to immunotherapy, and the outcome reported as the 1-year overall survival (OS). PD-L1 status (tumor proportion score ≥ 1%) was known for 70 cases. We preprocessed the H&amp;E images using two deep learning models previously developed using The Cancer Genome Atlas dataset. First, we used a classification model to identify tumor regions and randomly sampled a fixed number of tumor patches for each case. Then, we used a self-supervised pathology foundation model to obtain a compressed visual representation of each patch, known as an embedding. Next, using our dataset, we trained a deep multiple instance learning (DeepMIL) model with a gated attention mechanism to predict the binary 1-year OS status (0=deceased, 1=alive) for each case. As a baseline, we also trained a linear-probe (logistic regression) model using the averaged embeddings. Given the small dataset size, 5-fold cross-validation was used to train and evaluate both the DeepMIL and linear-probe models, with cases randomly split across folds. For evaluation, we used survival analysis to compare the 0/1 case groups. Overall, across all 150 cases, univariable Cox regression showed that 1-year OS was more strongly associated with the DeepMIL status (46/104, HR=0.55, p=0.03) than the linear-probe status (55/95, HR=0.81, p=0.44). Results were consistent on the subset of 70 cases with known PD-L1 status, whereby OS was most strongly associated with the DeepMIL status (19/51, HR=0.40, p=0.04) compared to the linear-probe status (31/39, HR=0.46, p=0.09) and PD-L1 status (30/40, HR=0.65, p=0.32). In multivariable Cox regression adjusting for age group and smoking status, OS remained more strongly associated with the DeepMIL status (HR=0.45, p=0.08) than PD-L1 status (HR=0.72, p=0.47). In conclusion, the DeepMIL status predicted from H&amp;E images showed a stronger association with outcomes compared to PD-L1 status, a standard biomarker for immunotherapy in NSCLC. These exploratory results demonstrate the potential of deep learning using pathology foundation models to improve immunotherapy outcomes prediction, even with small datasets. Such approaches may even enable the discovery of novel biomarkers from H&amp;E images to advance precision medicine. Citation Format: Jessica Loo, Yang Wang, Pok Fai Wong, Ellery Wulczyn, Jeremy Lai, Peter Cimermancic, David F. Steiner, Shamira S. Weaver. Predicting immunotherapy outcomes from H&amp;E images in lung cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 7380.

A differentiable first-order rule learner for inductive logic programming

Learning first-order logic programs from relational facts yields intuitive insights into the data. Inductive logic programming (ILP) models are effective in learning first-order logic programs from observed relational data. Symbolic ILP models support rule learning in a data-efficient manner. However, symbolic ILP models are not robust to learn from noisy data. Neuro-symbolic ILP models utilize neural networks to learn logic programs in a differentiable manner which improves the robustness of ILP models. However, most neuro-symbolic methods need a strong language bias to learn logic programs, which reduces the usability and flexibility of ILP models and limits the logic program formats. In addition, most neuro-symbolic ILP methods cannot learn logic programs effectively from both small-size datasets and large-size datasets such as knowledge graphs. In the paper, we introduce a novel differentiable ILP model called differentiable first-order rule learner (DFORL), which is scalable to learn rules from both smaller and larger datasets. Besides, DFORL only needs the number of variables in the learned logic programs as input. Hence, DFORL is easy to use and does not need a strong language bias. We demonstrate that DFORL can perform well on several standard ILP datasets, knowledge graphs, and probabilistic relation facts and outperform several well-known differentiable ILP models. Experimental results indicate that DFORL is a precise, robust, scalable, and computationally cheap differentiable ILP model.

Evolutionary neuron-level transfer learning for QoT estimation in optical networks

To ensure the reliability and efficiency of lightpath transmission in optical networks, it is essential to predict the quality of transmission (QoT) of the candidate lightpaths before their deployment. An artificial neural network (ANN) becomes a promising solution for QoT estimation of lightpaths, owing to its powerful data fitting capability and scalability. As a supervised learning model, an ANN requires a large set of training samples from lightpaths’ observations to ensure its accuracy. Unfortunately, the acquisition of the training samples is hindered by practical limitations, such as the shortage of monitorable lightpaths in the early stage of optical network deployment or/and the absence of optical performance monitors at partial network nodes. Therefore, how to obtain a high-precision QoT estimator with small-sized datasets is challenging. In this paper, we propose an evolutionary neuron-level transfer learning (ENTL) scheme for QoT estimation to improve the accuracy of the ANN model with small-sized datasets. In the ENTL-based QoT estimator, the minimal unit of knowledge transfer is the neuron of the ANN model, and the particle swarm optimization (PSO) algorithm is introduced to determine the trainable neurons and the frozen neurons, where the testing dataset generated by data augmentation assists the PSO algorithm to evaluate the feasible solutions. Simulation results show that the ENTL-based QoT estimator achieves higher accuracy than the traditional layer-level transfer learning (LTL)-based QoT estimator. And, when the ENTL-based QoT estimator is applied to optical network planning, it improves the reliability and throughput of optical networks compared with the LTL-based QoT estimator.

Military Unmanned Equipment Image Target Recognition Method based on Improved Deep Learning

Abstract Military unmanned equipment target recognition is currently a research hotspot and trend in the field of military intelligence. The small sample size and complex recognition scenarios of military unmanned equipment image datasets result in low recognition accuracy. It is proposed a military unmanned equipment image target recognition method based on improved deep learning, which set Faster R-CNN as the network framework for target recognition, used the Kmeans++ algorithm to label boxes on customized datasets, and then added OHEM to the framework to improve the network’s recognition accuracy for difficult to recognize samples. The accuracy of the algorithm proposed in this article reaches 93.8%, which is 2.8% higher than the YOLOv5 algorithm, providing an improved deep learning method for military unmanned equipment image target recognition.