No-reference image quality assessment is crucial for evaluating perceptual quality across diverse image-processing applications. Given the challenge of accruing mean opinion scores for images, utilizing data augmentation and transfer learning is vital for training predictive networks. This paper presents a new iterative transfer learning technique, which helps to transfer knowledge between heterogeneous network architectures, and overcomes the problem of overlearning when training on small datasets. The proposed method used a large amount of unlabeled data during training, improving its ability to handle different image quality conditions. We also presented a two-branch convolutional neural network architecture, which merges multi-scale and multi-level attributes efficiently. This architecture emphasizes both local detail extraction and high-level comprehension, and the result was fast execution time and minimal memory overhead. Empirical results showed that applying iterative transfer learning to train a two-branch convolutional neural network achieved superior real-time performance and at the same time exhibited good performance in spearman’s rank order correlation coefficient. Furthermore, the model manifested robustness for the noisy mean opinion score, which is prevalent in available datasets, and during data augmentation processes.

Biochemical oxygen demand (BOD) is an important water quality measurement but takes five days or more to obtain. This may result in delays in taking corrective action in water treatment. Our goal was to develop a BOD predictive model that uses other water quality measurements that are quicker than BOD to obtain; namely pH, temperature, nitrogen, conductivity, dissolved oxygen, fecal coliform, and total coliform. Principal component analysis showed that the data spread was in the direction of the BOD eigenvector. The vectors for pH, temperature, and fecal coliform contributed the greatest to data variation, and dissolved oxygen negatively correlated to BOD. K-means clustering suggested three clusters, and t-distributed stochastic neighbor embedding showed that BOD had a strong influence on variation in the data. Pearson correlation coefficients indicated that the strongest positive correlations were between BOD, and fecal and total coliform, as well as nitrogen. The largest negative correlation was between dissolved oxygen, and BOD. Multiple linear regression (MLR) using fecal, and total coliform, dissolved oxygen, and nitrogen to predict BOD, and training/test data of 80%/20% and 90%/10% had performance indices of RMSE=2.21 mg/L, r=0.48 and accuracy of 50.1%, and RMSE=2.18 mg/L, r=0.54 and an accuracy of 55.5%, respectively. BOD prediction was better than previous MLR models. Increasing the percentage of the training set above 80% improved the model accuracy but did not significantly impact its prediction. Thus, MLR can be used successfully to estimate BOD in water using other water quality measurements that are quicker to obtain.

The use of synthetic data could facilitate data-driven innovation across industries and applications. Synthetic data can be generated using a range of methods, from statistical modeling to machine learning and generative AI, resulting in datasets of different formats and utility. In the health sector, the use of synthetic data is often motivated by privacy concerns. As generative AI is becoming an everyday tool, there is a need for practice-oriented insights into the prospects and limitations of synthetic data, especially in the privacy sensitive domains. We present an interdisciplinary outlook on the topic, focusing on, but not limited to, the Finnish regulatory context. First, we emphasize the need for working definitions to avoid misplaced assumptions. Second, we consider use cases for synthetic data, viewing it as a helpful tool for experimentation, decision-making, and building data literacy. Yet the complementary uses of synthetic datasets should not diminish the continued efforts to collect and share high-quality real-world data. Third, we discuss how privacy-preserving synthetic datasets fall into the existing data protection frameworks. Neither the process of synthetic data generation nor synthetic datasets are automatically exempt from the regulatory obligations concerning personal data. Finally, we explore the future research directions for generating synthetic data and conclude by discussing potential future developments at the societal level.

Clustering is a fundamental technique for revealing structure within data. Density-based clustering has the advantage of detecting structure without prior knowledge, such as the distribution or the number of clusters. However, existing density-based clustering algorithms usually suffer from limitations in latent density patterns, which hinder their applicability to complex datasets. We have showed that this may be caused by the bias of standard methods for point density estimation, such as kernel density estimation (KDE), which tends to favor compact clusters. To address the limitations of existing methods, we introduced the concept of local comparative density. We then proposed a specialized technique well-suited for capturing the nature of local comparative density. This approach allowed us to balance the density variations across clusters. Building on this foundation, we presented a halo-based clustering method that can effectively discover clusters with arbitrary shapes, heterogeneous densities, peakless distributions, and unbalanced cluster sizes. Additionally, this method is capable of identifying high-quality noise within the data. The superiority of the proposed algorithm was demonstrated by comparing it with state-of-the-art density-based clustering algorithms on artificial and real-world datasets.

In recent years, transformer models have achieved great success in natural language processing (NLP) tasks. Most of the current results are achieved by using monolingual transformer models, where the model is pre-trained using a single-language unlabelled text corpus. Then, the model is fine-tuned to the specific downstream task. However, the cost of pre-training a new transformer model is high for most languages. In this work, we propose a cost-effective transfer learning method to adopt a strong source language model, trained from a large monolingual corpus to a low-resource language. Thus, using the XLNet language model, we demonstrate competitive performance with mBERT and a pre-trained target language model on the cross-lingual sentiment (CLS) dataset and on a new sentiment analysis dataset for the low-resource language Tigrinya. With only 10k examples of the given Tigrinya sentiment analysis dataset, English XLNet achieved 78.88% F1-Score, outperforming BERT and mBERT by 10% and 7%, respectively. More interestingly, fine-tuning (English) XLNet model on the CLS dataset showed promising results compared to mBERT and even outperformed mBERT for one dataset of the Japanese language.

In this paper, a physics-informed neural network model is proposed to predict the growth of online social network users. The number of online social network users is modeled by a stochastic process and the associated Kolmogorov forward equation is derived. Then, a physics-informed neural network model is built based on the Kolmogorov forward equation and trained using real-world data. By combining mathematical modeling with machine learning, our approach provides a practical and interpretable methodology that harnesses the strengths of both physical laws and advancements in machine learning, while minimizing the opacity in machine learning models.

Mopsi is a location-based platform for storing photos and GPS tracks of the users. It allowed user to share their data on-line with real-time user location, communicate with other users, share the data in Facebook, browse the collected photos and tracks on map, perform searches and ask recommendations. Mopsi was operational from 2009 to 2020. This paper documents the history of Mopsi, its main functionalities, research achievements, and the collected data.

Transformer-based models have shown to be highly effective for dealing with complex tasks in a wide range of areas due to their robust and flexible architecture. However, their generic nature frequently limits their effectiveness for domain-specific tasks unless significantly fine-tuned. We understand that behavioral health plays a vital role in individual well-being and community safety, as it influences interpersonal interactions and can significantly impact public safety. As a result, identifying and classifying these cases demands the use of an effective tool, such as a framework, that has been fine-tuned to context-specific behavioral health issues. In this work, we demonstrated a trainable lightweight approach for addressing behavioral health analysis utilizing feature embeddings generated from transformer-based models. To facilitate in domain adaptation, we created instruction sets based on annotations by subject matter experts, enabling for targeted fine-tuning of the large language model (LLM) for behavioral health applications. Our experiments demonstrated that parameter-frozen transformer-based models can capture high-quality feature representations that allowed for the integration of a lightweight framework, making them especially useful in resource-constrained settings.

Appearance defects frequently occur during cigarette manufacturing due to production equipment or raw materials. Appearance defects significantly impact the quality of tobacco products. Since manual inspection cannot keep pace with the demands of high-speed production lines, rapid and accurate automated classification and detection are essential. Supervised learning is predominantly employed in research on automated classification of product quality appearance defects. However, supervised learning necessitates substantial labeled data for training, which is time-consuming to annotate and prone to errors. This paper proposes a self-supervised learning-based classification model for cigarette appearance defects. This is a generative adversarial network (GAN) model based on masked autoencoders (MAE), called MAE-GAN. First, this model combines MAE as a generator with a simple discriminator to form a generative adversarial network according to the principle of mask reconstruction in MAE. The generator reconstructs the images to learn their features. Second, the model also integrates MAE’s loss function into the GAN’s loss function. This lets the model focus on pixel-level losses during training. As a result, model performance is improved. Third, a Wasserstein GAN with gradient penalty (WGAN-GP) is added to stabilize the training process. In addition, this paper preprocesses cigarette images through segmentation and recombination. Neural networks typically accept images with the same width and height. Due to the narrow shape of cigarette images, if the image is directly transformed into a square and fed into a neural network, the details of the image will be severely lost. This paper segments the cigarette image into three parts and recombines them into images with similar length and width, greatly improving classification accuracy.

Intelligent district heating control requires knowing the customers’ past behavior and predicting their future needs. This can reduce peak energy use, optimizing energy production, accurate billing, and reducing fraud. Clustering has been used for analyzing large-scale building operational data and recognizing consumption profiles. In this work, we analyze the heat consumption profiles of district heat customers in Kuopio, Finland. We constructed two consumption profiles of their average hourly use: one for weekdays, and one for weekends. Clustering is then used to construct four consumption profiles. These profiles can be used for intelligent control, prediction of future use, and to recognize abnormal use behavior. The latter can be the first indication of a problem like heat leaking, which can prevent possible water damage.

Autonomous vehicles (AV) face significant challenges in navigating adverse weather, particularly rain, due to the visual impairment of camera-based systems. In this study, we leveraged contemporary deep learning techniques to mitigate these challenges, aiming to develop a vision model that processes live vehicle camera feeds to eliminate rain-induced visual hindrances, yielding visuals closely resembling clear, rain-free scenes. Using the Car Learning to Act (CARLA) simulation environment, we generated a comprehensive dataset of clear and rainy images for model training and testing. In our model, we employed a classic encoder-decoder architecture with skip connections and concatenation operations. It was trained using novel batching schemes designed to effectively distinguish high-frequency rain patterns from low-frequency scene features across successive image frames. To evaluate the model’s performance, we integrated it with a steering module that processes front-view images as input. The results demonstrated notable improvements in steering accuracy, underscoring the model’s potential to enhance navigation safety and reliability in rainy weather conditions.

A long-range, wide-area network is a cost-effective, energy-efficient technology for wide-area sensor networks. But the massive Internet of Things (IoT) brings challenges such as increased traffic and energy consumption. Thus, there is a pressing need to design a scheduling strategy to improve network energy efficiency without compensating for its reliability. We have proposed a deep deterministic policy gradient-based scheduling algorithm with a frequency-hopping spread spectrum that avoids repeated collisions and retransmissions. Frequency-hopping divides frequency channels into subchannels, allowing multiple devices to operate simultaneously. This makes it a favorable scheduling strategy for dense networks, as it reduces collisions and energy consumption. Scheduling in a long-range, wide-area network involves selecting transmission parameters for each device, which can be cumbersome. We used the deep deterministic policy gradient algorithm to optimize schedule generation for high-density networks, enhancing energy efficiency. In this paper, we compared the performance of the frequency-hopping spread spectrum with other heuristic and machine learning-based algorithms using the LoRaSim simulator. We observed a 42% increase in the packet delivery ratio and a 17% improvement in energy efficiency with our solution, along with detailed results on the transmission time and collision reduction.

Large language models have shown strong capabilities in performing natural language planning tasks, largely due to the chain-of-thought method, which enhances their ability to solve complex tasks through explicit intermediate inference. However, they face challenges in acquiring new knowledge, executing calculations, and interacting with the environment. Although previous work has enabled large language models to use external tools to improve reasoning and environmental interaction, there was no scalable or cohesive structure for these technologies. In this paper, we present LLM-Collab, where Collab represents the cooperative interaction between two AI agents, and the large language model plays a key role in the creation of AI agents. For this method, we took large language models as the reasoning core for AI agents and designed two AI agents to cooperate on the planning tasks: One as an analyst for tool selection and phase validation, and the other as an executor of specific tasks. Our method provided a comprehensive list of external tools to facilitate the invocation and integration of agents, ensuring a seamless collaboration process. This paradigm established a unified framework for autonomous task-solving based on massive language models by demonstrating how language communication and tool selection enable multi-agent collaboration.

This paper presents a comprehensive study aimed at understanding the dynamics of single-vehicle and multi-vehicle crashes through a binary classification approach. By harnessing high-resolution, multi-source data, including high-resolution traffic profile data captured by weigh-in-motion stations, weather conditions, roadway attributes, and pavement properties, we delved into distinctive characteristics of the two crash types. Particularly, a meticulous data fusion approach was applied to integrate the diverse data sources, enabling a holistic investigation of influential factors. Framing it as a classification task, key factors differentiating between single-vehicle and multi-vehicle crashes were identified. The results of the study provide valuable insights into the underlying mechanisms of the two distinct crash types, supporting the development of targeted safety measures.