Synthetic Training Samples Research Articles

Exploring large language models for the generation of synthetic training samples for aspect-based sentiment analysis in low resource settings

Aspect-Based Sentiment Analysis (ABSA) is a fine-grained task in sentiment analysis, aiming to identify sentiment expressed towards specific aspects of an entity. This paper explores the use of Large Language Models (LLMs), specifically GPT-3.5-turbo and Llama-3-70B, for generating annotated data in Aspect-Based Sentiment Analysis (ABSA), aiming to address the scarcity of labeled datasets in the field. Two low-resource scenarios are considered, with 25 and 500 manually annotated examples available. In the 25-example scenario, adding synthetic examples generated through few-shot prompting resulted in F1 scores of 81.33 for Aspect Category Detection (ACD) and 71.71 for Aspect Category Sentiment Analysis (ACSA). For the 500-example scenario, synthetic data augmentation showed a notable gain only for the ACSA task, raising the F1 score from 84.54 to 86.70.

Non-Fluent Synthetic Target-Language Data Improve Neural Machine Translation.

When the amount of parallel sentences available to train a neural machine translation is scarce, a common practice is to generate new synthetic training samples from them. A number of approaches have been proposed to produce synthetic parallel sentences that are similar to those in the parallel data available. These approaches work under the assumption that non-fluent target-side synthetic training samples can be harmful and may deteriorate translation performance. Even so, in this paper we demonstrate that synthetic training samples with non-fluent target sentences can improve translation performance if they are used in a multilingual machine translation framework as if they were sentences in another language. We conducted experiments on ten low-resource and four high-resource translation tasks and found out that this simple approach consistently improves translation performance as compared to state-of-the-art methods for generating synthetic training samples similar to those found in corpora. Furthermore, this improvement is independent of the size of the original training corpus, the resulting systems are much more robust against domain shift and produce less hallucinations.

Synthetic Data and Hierarchical Object Detection in Overhead Imagery

The performance of neural network models is often limited by the availability of big datasets. To treat this problem, we survey and develop novel synthetic data generation and augmentation techniques for enhancing low/zero-sample learning in satellite imagery. In addition to extending synthetic data generation approaches, we propose a hierarchical detection architecture to improve the utility of synthetic training samples. We consider existing techniques for producing synthetic imagery–3D models and neural style transfer–as well as introducing our own adversarially trained reskinning network, the GAN-Reskinner, to blend 3D models. Additionally, we test the value of synthetic data in a two-stage, hierarchical detection/classification model of our own construction. To test the effectiveness of synthetic imagery, we employ it in the training of detection models and our two stage model, and evaluate the resulting models on real satellite images. All modalities of synthetic data are tested extensively on practical, geospatial analysis problems. Our experiments show that synthetic data developed using our approach can often enhance detection performance, particularly when combined with some real training images. When the only source of data is synthetic, our GAN-Reskinner often boosts performance over conventionally rendered 3D models and in all cases, the hierarchical model outperforms the baseline end-to-end detection architecture.

AdSMOTE: A Technique for “High Proportion” Audio Augmentation

Data augmentation is a practice that is widely used in the fields of machine and deep learning. It is used primarily for its effectiveness in reducing the generalisation gap between training and validation, as well as to artificially increase in available training data points. This is particularly relevant to audio datasets, which are usually smaller and suffer from imbalanced classes in some applications. This work presents adSMOTE (audio SMOTE), a novel sampling and augmentation strategy and also compares it to Specaugment, one of the most effective augmentation strategies for audio data. We show that our method outperforms the latter by a considerable margin when the proportion of synthetic training samples is high. We also provide source code for the complete algorithm, which can easily be integrated into an existing model, enabling the rapid development of augmentation frameworks.

A balancing act: Synthetic oversampling corrects skewed distribution of aflatoxin in maize and improves prediction of contaminated kernels

Current spectral models for detecting aflatoxin (AF) in maize kernels are limited by very low representation of kernels with AF levels above the models’ specified contamination thresholds. With the aim of improving prediction model sensitivity to AF-contaminated kernels, two methods were evaluated for artificially enriching the representation of kernels containing ≥150 parts per billion (ppb) AF in a large spectral dataset: humid incubation and synthetic oversampling. The addition of synthetic training samples improved prediction accuracy of kernels containing ≥150 ppb AF from 51% to 80%. Humid incubation contributed additional kernels with intermediate levels of AF contamination (primarily 5–75 ppb), but did not change the overall distribution of AF contamination. Feature importance distributions overlapped among models at 329–345 nm, 380–385.5 nm, 415–425 nm, 639–668 nm, and 1013.5–1060 nm. These spectral ranges could be applicable to the development of limited wavelength grain sorting devices for low-resource settings.

Conditional synthetic food image generation

Generative Adversarial Networks (GAN) have been widely investigated for image synthesis based on their powerful representation learning ability. In this work, we explore the StyleGAN and its application of synthetic food image generation. Despite the impressive performance of GAN for natural image generation, food images suffer from high intra-class diversity and inter-class similarity, resulting in overfitting and visual artifacts for synthetic images. Therefore, we aim to explore the capability and improve the performance of GAN methods for food image generation. Specifically, we first choose StyleGAN3 as the baseline method to generate synthetic food images and analyze the performance. Then, we identify two issues that can cause performance degradation on food images during the training phase: (1) inter-class feature entanglement during multi-food classes training and (2) loss of high-resolution detail during image downsampling. To address both issues, we propose to train one food category at a time to avoid feature entanglement and leverage image patches cropped from high-resolution datasets to retain fine details. We evaluate our method on the Food-101 dataset and show improved quality of generated synthetic food images compared with the baseline. Finally, we demonstrate the great potential of improving the performance of downstream tasks, such as food image classification by including high-quality synthetic training samples in the data augmentation.

Advantage of Machine Learning over Maximum Likelihood in Limited-Angle Low-Photon X-Ray Tomography

Limited-angle X-ray tomography reconstruction is an ill-conditioned inverse problem in general. Especially when the projection angles are limited and the measurements are taken in a photon-limited condition, reconstructions from classical algorithms such as filtered backprojection may lose fidelity and acquire artifacts due to the missing-cone problem. To obtain satisfactory reconstruction results, prior assumptions, such as total variation minimization and nonlocal image similarity, are usually incorporated within the reconstruction algorithm. In this work, we introduce deep neural networks to determine and apply a prior distribution in the reconstruction process. Our neural networks learn the prior directly from synthetic training samples. The neural nets thus obtain a prior distribution that is specific to the class of objects we are interested in reconstructing. In particular, we used deep generative models with 3D convolutional layers and 3D attention layers which are trained on 3D synthetic integrated circuit (IC) data from a model dubbed CircuitFaker. We demonstrate that, when the projection angles and photon budgets are limited, the priors from our deep generative models can dramatically improve the IC reconstruction quality on synthetic data compared with maximum likelihood estimation. Training the deep generative models with synthetic IC data from CircuitFaker illustrates the capabilities of the learned prior from machine learning. We expect that if the process were reproduced with experimental data, the advantage of the machine learning would persist. The advantages of machine learning in limited angle X-ray tomography may further enable applications in low-photon nanoscale imaging.

A Generative Approach to Open Set Recognition Using Distance-Based Probabilistic Anomaly Augmentation

Machine learning (ML) algorithms that are used in decision support (DS) and autonomous systems commonly train on labeled categorical samples from a closed set. This, however, poses a problem for deployed DS and autonomous systems when they encounter an anomalous pattern that did not originate from the closed set distribution used for training. In this case, the ML algorithm that was trained only on closed set samples may erroneously identify an anomalous pattern as having originated from one of the categories in the closed set, sometimes with very high confidence. In this paper, we consider the problem of unknown pattern recognition from a generative perspective in which additional synthetic training samples that represent anomalies are added to the training data. These synthetic samples are generated to optimally balance the desire to place anomalies all along the boundary of the training set in feature space, while not adversely effecting core classification performance on the test set. We demonstrate the efficacy of distance-based probabilistic anomaly augmentation (DPAA) that is proposed in this paper for a diverse set of applications such as character recognition and intrusion detection, and compare its combined classification and identification performance to both recent open set and more traditional novelty detection approaches.

Data Fusion for Environmental Process Control: Maximizing Useful Information Recovery under Data Limited Constraints

Real-time sensing of minor, or difficult to measure, components in complex systems is a challenge faced across disciplines. For environmental applications, this often entails measurement of target chemicals-that may be harmful and/or of interest at very low concentrations - in a complex medium (tens to thousands of constituent components). Sensor arrays and machine learning (ML) approaches have demonstrated some success but remain limited by the application context and high labor costs of chemical assays for creating large training datasets. This article explores a lower overhead approach, employing data fusion techniques to extract information from a relatively small training dataset. Creation of statistically relevant synthetic training samples is utilized to reduce dependence on costly analysis of samples from the target system. Samples were characterized using eight sensor modalities to create a training set for several ML algorithms, namely artificial neural networks (ANN), support vector regression (SVR), and random forests (RF), to measure NH 4 + concentrations (≤50 μM) online in a game-changing wastewater treatment process. Hyperparameters for each method were tuned using a particle swarm optimization approach, and both the accuracy and consistency of results were evaluated. ANN achieved the lowest mean absolute error (MAE) of ~6 μM, but all methods had a minimum MAE within 20% of this value. When evaluated on computational demand, SVR outperformed other approaches. ANN and RF showed wide variation in resulting MAE for a given parameterization, demonstrating strong dependence on initialization and training process. Overall RF provided the best balance of accuracy and consistency, and therefore, in applications where data are expected to be updated frequently or computational resources are not infinite, RF may provide the best tradeoff in speed, accuracy, and consistency.

Ship detection in spaceborne infrared images based on Convolutional Neural Networks and synthetic targets

Automatic detection of ships in spaceborne infrared images is important for both military and civil applications due to its all-weather detection capability. While deep learning methods have made great success in many image processing fields recently, training deep learning models still relies on large amount of labeled data, which may limit its application performance for infrared images target detection tasks. Considering that, we propose a new infrared ship detection method based on Convolutional Neural Networks (CNN) which is trained only with synthetic targets. For the problem of limited infrared training data, we design a Transfer Network (T-Net) to generate large amount of synthetic infrared-style ship targets from Google Earth images. The experiments are conducted on a near infrared band image (0.845–0.885μm), a short wavelength infrared band image (1.560–1.66μm) and a long wavelength infrared band image (2.1–2.3μm) of Landsat-8 satellite. The results demonstrate the effectiveness of the target generation ability of T-Net. With only synthetic training samples, our detection method achieves a higher accuracy than other classical ship detection methods.

Synthetic training samples for enhanced locality‐constrained dictionary learning

Dictionary learning serves as a considerable role in image processing and pattern recognition. However, when applied to face classification, it may suffer from the issue of the limited quantity of training samples. Therefore, it becomes a challenge to obtain a robust and discriminative dictionary. Recently, locality-constrained and label embedding dictionary learning (LCLE-DL) takes the locality and label information of atoms into account to achieve an effective performance in image classification. In this study, the authors exploit a new approach which uses synthetic training samples to enhance this dictionary learning algorithm, so they name it STS-DL. Firstly, they strengthen the diversities of training samples by producing virtual samples. Secondly, the LCLE-DL algorithm is used to calculate two deviations on the basis of the original training samples and the authors’ newly synthetic samples, respectively. Finally, they integrate them together to perform the classification task, which produces a more promising performance for image recognition. Abundant experiments have been conducted on several benchmark databases, the experimental results illustrate that the proposed STS-DL shows a higher accuracy than the LCLE-DL method, as well as some state-of-the-art dictionary learning and sparse representation algorithms in image classification.

An Immune Clonal Selection Algorithm for Synthetic Signature Generation

The collection of signature data for system development and evaluation generally requires significant time and effort. To overcome this problem, this paper proposes a detector generation based clonal selection algorithm for synthetic signature set generation. The goal of synthetic signature generation is to improve the performance of signature verification by providing more training samples. Our method uses the clonal selection algorithm to maintain the diversity of the overall set and avoid sparse feature distribution. The algorithm firstly generates detectors with a segmentedr-continuous bits matching rule andP-receptor editing strategy to provide a more wider search space. Then the clonal selection algorithm is used to expand and optimize the overall signature set. We demonstrate the effectiveness of our clonal selection algorithm, and the experiments show that adding the synthetic training samples can improve the performance of signature verification.