Margin-Aware Preference Optimization for Aligning Diffusion Models Without Reference

With the increasing integration of artificial intelligence (AI) in several scientific domains, there is a rising demand for advanced AI tools capable of addressing advanced research challenges. A challenge of paramount importance lies in accurately predicting the streamflow within river basins. Effective river flow prediction holds significant relevance, particularly given the substantial societal implications of river usage, encompassing areas such as transportation, agriculture, and power generation. The present study introduces a novel approach to streamflow prediction involving the development of a Deep Learning (DL) model that combines a convolutional neural network with Transfer Learning (TL) techniques to predict streamflow in river systems. With the aim of training the developed DL model, the study employed a time-series dataset containing hydrological data related to two distinct river basins, i.e., Paraíba do Sul, in Brazil, and Zambezi in the state of Mozambique. The developed DL models exhibited the capability to effectively predict the river flow with a one-day horizon, relying on the preceding three or seven days of historical data. To overcome the limited availability of training data and reduce the training time of DL models, TL was leveraged to incorporate two additional distinct time-series datasets, i.e., historical streamflow data from the São Francisco River in Brazil, and climate data from Delhi, India. The application of TL significantly reduced training time, leading only to a minimal decrease in prediction performance. Indeed, in the case DL models were trained on data collected from the Paraíba do Sul River, a substantial reduction in training time was observed - up to 27% - with a modest percentage decrease of 0.31% in test predictive performance (). Similarly, TL induced a significant reduction in training time of up to 48%, while resulting in a modest 2% reduction in test predictive performance () for the Zambezi dataset. The findings underscore the significance of TL as a strategic and viable approach to improve the efficiency of river flow prediction models in the context of basins with limited hydrological data available.

Owing to the unrestricted nature of the content in the training data, large text-to-image diffusion models, such as Stable Diffusion (SD), are capable of generating images with potentially copyrighted or dangerous content based on corresponding textual concepts information. This includes specific intellectual property (IP), human faces, and various artistic styles. However, Negative Prompt, a widely used method for content removal, frequently fails to conceal this content due to inherent limitations in its inference logic. In this work, we propose a novel strategy named Degeneration-Tuning (DT) to shield contents of unwanted concepts from SD weights. By utilizing Scrambled Grid to reconstruct the correlation between undesired concepts and their corresponding image domain, we guide SD to generate meaningless content when such textual concepts are provided as input. As this adaptation occurs at the level of the model's weights, the SD, after DT, can be grafted onto other conditional diffusion frameworks like ControlNet to shield unwanted concepts. In addition to qualitatively showcasing the effectiveness of our DT method in protecting various types of concepts, a quantitative comparison of the SD before and after DT indicates that the DT method does not significantly impact the generative quality of other contents. The FID and IS scores of the model on COCO-30K exhibit only minor changes after DT, shifting from 12.61 and 39.20 to 13.04 and 38.25, respectively, which clearly outperforms the previous methods.

This study focuses on harnessing deep-learning-based text-to-image transformation techniques to help webtoon creators’ creative outputs. We converted publicly available datasets (e.g., MSCOCO) into a multimodal webtoon dataset using CartoonGAN. First, the dataset was leveraged for training contrastive language image pre-training (CLIP), a model composed of multi-lingual BERT and a Vision Transformer that learnt to associate text with images. Second, a pre-trained diffusion model was employed to generate webtoons through text and text-similar image input. The webtoon dataset comprised treatments (i.e., textual descriptions) paired with their corresponding webtoon illustrations. CLIP (operating through contrastive learning) extracted features from different data modalities and aligned similar data more closely within the same feature space while pushing dissimilar data apart. This model learnt the relationships between various modalities in multimodal data. To generate webtoons using the diffusion model, the process involved providing the CLIP features of the desired webtoon’s text with those of the most text-similar image to a pre-trained diffusion model. Experiments were conducted using both single- and continuous-text inputs to generate webtoons. In the experiments, both single-text and continuous-text inputs were used to generate webtoons, and the results showed an inception score of 7.14 when using continuous-text inputs. The text-to-image technology developed here could streamline the webtoon creation process for artists by enabling the efficient generation of webtoons based on the provided text. However, the current inability to generate webtoons from multiple sentences or images while maintaining a consistent artistic style was noted. Therefore, further research is imperative to develop a text-to-image model capable of handling multi-sentence and -lingual input while ensuring coherence in the artistic style across the generated webtoon images.

Currently, generative artificial intelligence technologies are being actively developed and widely implemented. These technologies are used to solve various tasks in many areas of human activity, including image stylization in digital art, design, advertising, entertainment, and other fields. Image stylization consists in the transformation of its content while maintaining the semantic structure, but with a change in the visual style set by the sample or model parameters. Various methods are used to generate images, of which neural style transfer, variational autoencoders, generative-adversarial networks and diffusion models are considered in the article. The results of their analysis showed that the best approach for creating artistically stylized images is the use of generative-adversarial networks. There are a number of such network architectures, among which Pix2Pix, CycleGAN, StyleGAN, StarGANv and UGATIT have been investigated. Their comparative analysis showed that the CycleGAN architecture best meets the requirements of the task of stylizing images in the style of a particular painter. It does not require paired data, demonstrates resilience to different image structures, and provides stable and reproducible learning. Taken together, these advantages make CycleGAN a preferred choice for building an artistic stylization system. To conduct the experiments, an image stylization system was developed, trained on a set of images with copies of artworks by Vincent Van Gogh, Claude Monet, Henri Matisse, Ivan Aivazovsky and Ilya Repin. These authors were selected based on a set of criteria: popularity, recognition of style, genre diversity and the availability of a sufficient number of digital copies of artworks. Metrics were used to evaluate the quality of image generation. To evaluate the quality of generated images, metrics such as Kernel Inception Distance, Inception Score, and Contrastive Language-Image Pretraining based cosine similarity were used. The experimental results showed that generative models demonstrate the potential to mimic common features of artistic styles. However, in order to achieve a higher level of reliability and accuracy, further research is needed to adapt them to the given styles.

Parallel data mining algorithms are extensively used to mine and discover hidden knowledge from varied, unrelated data. Parallel data mining algorithms provide advantages such as reduced training time, less execution time, and less memory requirement. There are several issues in executing parallel data mining algorithms in a distributed environment. It is crucial to partition the data among processors such that there is minimal data dependency, proper synchronization, communication overhead, work load balancing among nodes in distributed processors and disk IO cost. Few of these issues can be resolved when parallel data mining algorithms are executed on Apache framework called Hadoop Map Reduce. Hadoop Map Reduce provides improved performance, reduced communication cost, reduced execution time, reduced training time, and reduced IO access. This paper proposes a novel framework that aims at enhancing the aforementioned advantages in terms of scalability by increasing the number of nodes in the Hadoop cluster and analyzing the performance of classification algorithms like K-Nearest Neighbor, Naive Bayes and Decision Tree. This parallel framework could be extended to other fields of biotechnology where prediction on large datasets is essential.

Fine-tuning of large language models is often demanding in terms of computational resources and memory. Consequently, there is a need to explore new methods that can effectively fine-tune these models without compromising performance. In this paper, we compared traditional and parameter-efficient fine-tuning of large language model XLM-RoBERTa, across both high-resourced languages like English and low-resourced languages such as Serbian. The model is fine-tuned for sentiment analysis when it is conducted using text classification. The experiments have shown that parameter-efficient fine-tuning can produce results that are on par with traditional fine-tuning, both for the English and the Serbian language. Furthermore, parameter efficient fine-tuning approach significantly reduces the computational time and memory requirements associated with model training, thereby offering a promising avenue for resource-efficient large language models fine-tuning. In experiments with the English language, we achieved accuracy using the PEFT approach that was only 3% lower than the accuracy obtained through the traditional approach, while reducing training time by 35% and memory requirements by 42%. In experiments involving the Serbian language, we achieved accuracy using the PEFT approach that was only 1% lower than the accuracy obtained through the traditional approach, while reducing training time by 42% and memory requirements by 69%.

Distributed machine learning (ML) can bring more computational resources to bear than single-machine learning, thus enabling reductions in training time. Distributed learning partitions models and data over many machines, allowing model and dataset sizes beyond the available compute power and memory of a single machine. In practice though, distributed ML is challenging when distribution is mandatory, rather than chosen by the practitioner. In such scenarios, data could unavoidably be separated among workers due to limited memory capacity per worker or even because of data privacy issues. There, existing distributed methods will utterly fail due to dominant transfer costs across workers, or do not even apply. We propose a new approach to distributed fully connected neural network learning, called independent subnet training (IST), to handle these cases. In IST, the original network is decomposed into a set of narrow subnetworks with the same depth. These subnetworks are then trained locally before parameters are exchanged to produce new subnets and the training cycle repeats. Such a naturally "model parallel" approach limits memory usage by storing only a portion of network parameters on each device. Additionally, no requirements exist for sharing data between workers (i.e., subnet training is local and independent) and communication volume and frequency are reduced by decomposing the original network into independent subnets. These properties of IST can cope with issues due to distributed data, slow interconnects, or limited device memory, making IST a suitable approach for cases of mandatory distribution. We show experimentally that IST results in training times that are much lower than common distributed learning approaches.

• Propose a novel generative framework FAFGAN for fuel consumption estimation. • FAFGAN estimates road-level fuel consumption using only macroscopic traffic observations. • A differentiable fuzzy logic module is developed to infer driving style for fuel estimation. • FAFGAN outperforms classic methods with more than 70% reduction in training time. For estimating the road-level fuel consumption, the existing models rely on data collected from high-precision detectors and rarely incorporate behavioral factors such as the impact of driving styles, which limits the accuracy and the scalability of the estimation models for real-world applications. To this end, this paper proposes a novel generative framework that estimates fuel consumption at the road level with low cost and high behavioral consistency, using only macroscopic traffic observations. The model consists of three key modules: a speed generation module conditioned on macroscopic inputs to reconstruct vehicle trajectory data, a differentiable fuzzy logic module to capture the uncertainty and heterogeneity of driving styles, and a style-aware fusion module that embeds behavioral semantics into the fuel estimation process. We tested the proposed model against the baseline models on the NGSIM and HighD datasets. The results show that FAFGAN achieves the best road-level estimation accuracy with RMS E avg , outperforming advanced generative methods including LCM, HRF, and SiT by 3% to 14%, while achieving comparable accuracy to GaGAN with 72% less training time.

A vision-based traffic surveillance processor is being developed at JPL. This processor uses innovative image segmentation and classification techniques for vehicles in freeway images, including those with large shadows. These results enable the computation of many useful traffic parameters. A wavelet-based algorithm has been developed for vehicle detection and segmentation. Specifically, two types of mother wavelet has been created and tested: the first for shape-size discrimination of vehicles from their background; and the second for locating where vehicles join their shadows, thus enabling segmentation of the vehicles from their shadows. Combining these two wavelets enables robust segmentation of vehicles from busy freeways. This method reduces the false-alarm rate in vehicle counts, since shadows are no longer mistaken for vehicles. We use neural networks for vehicle classification. To reduce system complexity and training time, we use, as preprocessors, several feature extraction methods, such as invariant-moment and Hermite-moment computations. This preprocessing enables orders of magnitude reductions in training time and a great increase in classification accuracy.

This study quantifies how systematic feature reduction affects detection quality and computational cost for three tree-based classifiers (Decision Tree, Random Forest, and XGBoost) on the NSL-KDD intrusion-detection benchmark. Principal Component Analysis and SHAP-based feature ranking were applied to produce subsets of 30, 20, 15, 10, 5, and 3 features from the original 41. For each subset, models were optimized using Optuna and evaluated on training time, prediction latency, accuracy, and weighted F1 score. Results indicate that XGBoost sustains the greatest robustness under aggressive reduction: an F1 of 0.9962 with all features and 0.9700 with only three features, while achieving approximately a 70 percent reduction in training time at extreme reduction levels. Decision Tree exhibited the lowest inference latency but experienced larger performance loss below ten features. Random Forest maintained competitive F1 values but incurred substantially higher training overhead at moderate to high dimensions. These findings provide guidance for deploying intrusion-detection models in resource-constrained settings, highlighting trade-offs between detection fidelity and computational burden and informing choices for real-time monitoring applications.

Decision Making in Recurrent Neuronal Circuits

We introduce a generalized representation for a boosted classifier with multiple exit nodes, and propose a method to training which combines the idea of propagating scores across boosted classifiers [14, 17] and the use of asymmetric goals [13]. A means for determining the ideal constant asymmetric goal is provided, which is theoretically justified under a conservative bound on the ROC operating point target and empirically near-optimal under the exact bound. Moreover, our method automatically minimizes the number of weak classifiers, avoiding the need to retrain a boosted classifier multiple times for empirical best performance as in conventional methods. Experimental results shows significant reduction in training time and number of weak classifiers, as well as better accuracy, compared to conventional cascades and multi-exit boosted classifiers.

Online signature verification is one of the biometric authentication techniques. Online signature verification reduces the human error that may occur due to physical signature verification. Many researchers have provided a method of online signature verification still there is a scope for improvement. In this paper, a novel blended feature vector (BFV) is formed by combining two feature vectors. The first of these feature vectors is formed using a raw online signature database and the other feature vector is formed of the features extracted from the grayscale images obtained from the signature database. The algorithm makes use of the probability distribution of the features to provide a feature vector that results in an enhanced verification system. Two online signature databases, namely SVC2004 and ATVS-SSig are used. BFV is used to train the proposed mixed sequence deep neural network (MS-DNN). The use of a blended feature vector in the training of MS-DNN, instead of a whole online signature database, improves the training time of deep neural network to a great extent. This reduction in training time might be useful for new user registration. The performance parameters considered are validation efficiency and equal error rate. Results are compared with the existing state of technology and the comparison shows that the use of a blended feature vector as a classification vector improves the validation efficiency (99.5%) and also there is an improvement in the verification success rate in terms of equal error rate (EER 1.5%) as compared to the existing research.

In this paper, we propose a symmetric difference data enhancement physics-informed neural network (SDE-PINN) to study soliton solutions for discrete nonlinear lattice equations (NLEs). By considering known and unknown symmetric points, numerical simulations are conducted to one-soliton and two-soliton solutions of a discrete KdV equation, as well as a one-soliton solution of a discrete Toda lattice equation. Compared with the existing discrete deep learning approach, the numerical results reveal that within the specified spatiotemporal domain, the prediction accuracy by SDE-PINN is excellent regardless of the interior or extrapolation prediction, with a significant reduction in training time. The proposed data enhancement technique and symmetric structure development provides a new perspective for the deep learning approach to solve discrete NLEs. The newly proposed SDE-PINN can also be applied to solve continuous nonlinear equations and other discrete NLEs numerically.

The superior soft-tissue contrast achieved using magnetic resonance imaging (MRI) compared to x-ray computed tomography (CT) has led to the popularization of MRI-guided radiation therapy (MR-IGRT), especially in recent years with the advent of first and second generation MRI-based therapy delivery systems for MR-IGRT. The expanding use of these systems is driving interest in MRI-only RT workflows in which MRI is the sole imaging modality used for treatment planning and dose calculations. To enable such a workflow, synthetic CT (sCT) data must be generated based on a patient's MRI data so that dose calculations may be performed using the electron density information derived from CT images. In this study, we propose a novel deep spatial pyramid convolutional framework for the MRI-to-CT image-to-image translation task and compare its performance to the well established U-Net architecture in a generative adversarial network (GAN) framework. Our proposed framework utilizes atrous convolution in a method named atrous spatial pyramid pooling (ASPP) to significantly reduce the total number of parameters required to describe the model while effectively capturing rich, multi-scale structural information in a manner that is not possible in the conventional framework. The proposed framework consists of a generative model composed of stacked encoders and decoders separated by the ASPP module, where atrous convolution is applied at increasing rates in parallel to encode large-scale features. The performance of the proposed method is compared to that of the conventional GAN framework in terms of the time required to train the model and the image quality of the generated sCT as measured by the root mean square error (RMSE), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR) depending on the size of the training data set. Dose calculations based on sCT data generated using the proposed architecture are also compared to clinical plans to evaluate the dosimetric accuracy of the method. Significant reductions in training time and improvements in image quality are observed at every training data set size when the proposed framework is adopted instead of the conventional framework. Over 1042 test images, values of 17.7±4.3HU, 0.9995±0.0003, and 71.7±2.3 are observed for the RMSE, SSIM, and PSNR metrics, respectively. Dose distributions calculated based on sCT data generated using the proposed framework demonstrate passing rates equal to or greater than 98% using the 3D gamma index with a 2%/2mm criterion. The deep spatial pyramid convolutional framework proposed here demonstrates improved performance compared to the conventional GAN framework that has been applied to the image-to-image translation task of sCT generation. Adopting the method is a first step toward an MRI-only RT workflow that enables widespread clinical applications for MR-IGRT including online adaptive therapy.