Generative Adversarial Networks (GANs) have become pivotal for generating synthetic data. This paper conducts a comprehensive comparison of three cutting-edge GAN models. In particular, this study delved deep into the architectural intricacies, strengths, and limitations of each model, emphasizing their distinct features and mechanisms. DF-GANs focus on producing natural images with a single-stage backbone, DM-GANs leverage memory structures to enhance model performance, while AttnGAN employs attention-driven, multi-stage refinement for precise text-to-image generation. Through a series of literature search, this study evaluates the applicability of these models in various scenarios, offering insights into their practical implications and potential areas of improvement. This comparative study aims to serve as a reference point for researchers and practitioners alike, shedding light on the contemporary advancements in GAN technology and guiding future developments in the domain.

Recent research attention has been captivated by the advent of Generative Adversarial Networks (GANs) in the realm of generating visuals from textual descriptions. Within a GAN framework, the interplay between the discriminator and generator components facilitates the production of lifelike visuals. This method proves to be versatile and user-friendly, allowing for the generation of authentic, diverse, and semantically faithful conditional images. However, the field still has to solve two issues: the development of high-resolution images with multiple elements and the construction of proper evaluation criteria that correlate with human perception. This paper contextualizes a number of adversarial text-to-image generation models and their core principles. This article engages in a comprehensive examination of the current methodologies employed in the analysis of text-to-image generation models, emphasizing their limitations and proposing avenues for future advancements. The discussion within this article centers on the utilization of generative adversarial networks in text-to-image synthesis, offering researchers both a comparative analysis and a benchmark for their text-to-image generation studies.

Gaussian-Process Factor Analysis (GPFA) is a useful method to discover the unknown dynamics of neural activities. Currently, there are a lot of studies based on the GPFA model. However, many of the existing GPFA models are specially for a specific situation, and they are no longer effective in other conditions. This paper aims to solve this problem by proposing a GPFA framework based on the standard GPFA model which can be applied to any neural dynamics with unknown latent structure. This framework also provides an idea to determine the latent dimension by using cross-validation. This framework will first be used on the synthetic data created by a generative model, to test two different ways of reproducing the spike train and check its utility. After that, it will be applied to real neural data recorded from anesthetized macaque monkeys. The framework shows a good result on the synthetic data. And its performance on the real neural data suggests that it still has some space to be improved. Discussion of the result will mainly focus on the potential approach to improve the framework’s accuracy.

As a matter of fact, consumer behaviour analysis plays a key role in sales strategy designs in recent years. Thanks to the rapid development of data analysis techniques as well as machine learning schemes, it is available to analyse the data collected in daily life to realize the consumer behaviour analysis (e.g., prediction, judgement). With this in mind, this study will select some of the specific cases with different models to discuss the implementation way. According to the analysis, the performances for the different models in different situations are discussed. At the same time, the current limitations for different applications as well as analysis pattern are clarified as well as evaluated based on the analysis. In the meantime, the prospects for the further study are demonstrated. Overall, these results shed light on guiding further exploration of implementation the big data analysis techniques as well as machine learning scenarios in consumer behaviour.

Cryptocurrency has caught huge amounts of investment attentions and interests ever since its first introduction. However, due to the highly instable nature of its price, it is crucial for investors to avoid risks when investing in cryptocurrency. Studies has been conducted to predict the price of cryptocurrency with different price influencing factors using various machine learning models. On the other hand, most of them only focused on major types of cryptocurrencies, e.g., Bitcoin and overlooked minor ones. This study focuses on one type of minor cryptocurrency, Litecoin. Three machine learning algorithms, random forest regression, light gradient boosting machine (LightGBM), and long short-term memory (LSTM) are used to predict long-term Litecoin price. The effects of 19 price influencing factors are considered, including Litecoin price variables, other popular cryptocurrency prices, major foreign exchange prices, market indices, and major commodities prices. Root mean squared error (RMSE), mean absolute percentage error (MAPE), and R-squared score are used to evaluate model performances. The results suggest that, among the 3 models, random forest model shows the best prediction with the least error, while LSTM model has the most error. Such result can provide insights for investors to avoid risks in Litecoin investments. Future studies are still necessary to take more types of cryptocurrency and price influencing factors into consideration.

In the swiftly advancing realm of Artificial Intelligence (AI), there has been a remarkable evolution in Natural Language Processing (NLP). Language models like ChatGPT have ushered in a new era of human-computer interaction, bestowing upon people an unparalleled capacity to comprehend and generate text that mimics human language. When these models are fine-tuned for specific domains, such as medicine, they become endowed with domain-specific knowledge, bridging the chasm between language comprehension and specialized expertise. However, a noteworthy research gap exists when it comes to harnessing AI-powered NLP for medical diagnosis and symptom analysis, particularly in the development of a Symptom to Disease Chatbot. This innovative approach aims to provide accurate and accessible healthcare information, enhancing the initial steps of medical consultation. This research proposes and develops a pioneering Symptom to Disease Chatbot, powered by the sophisticated Llama 2 13b pre-trained language model. By integrating extensive medical knowledge and AI capabilities, this chatbot seeks to empower individuals to make informed decisions about their health, potentially transforming personal health management. The research includes data preparation involving curated medical datasets, transforming them into a conversational format suitable for training. The base model, Llama 2, is fine-tuned for this medical context. While preliminary results are promising, further exploration with larger datasets is essential to enhance performance. Additionally, an open-source Gradio interface enhances user interaction. This research addresses the critical need for accessible healthcare information and demonstrates the potential of AI-powered language models in the healthcare sector.

The real estate sector is pivotal to economic growth and plays a substantial role in the global GDP. In this technologically advanced age, the adoption of machine learning for accurate house price prediction is crucial. These models optimize decision-making for homeowners, sellers, and investors alike. This study represents a comprehensive exploration into the field of house price prediction within the context of Ames, Iowa, United States. The primary objective of this research is to construct a reliable and highly accurate predictive model, empowering individuals to estimate property values with unprecedented precision. The research encompasses three different machine learning algorithms: linear regression, random forest, and XGBoost. The use of the dataset from the reputed website helps improve the reliability of the result. Furthermore, the investigation extends to a detailed examination of the multifaceted determinants exerting a profound influence on house prices in the dynamic Ames real estate landscape, and determined that the factor that will influence the house price most is the total area of the house. Among all models, XGBoost produces the best result, which achieved an R-square of 0.8803. Moreover, the importance of each feature is also analyized using the feature ranking algorithm in random forest, showing that the overall quantity of the house, the living area of the house, and the total area of the basement are the top three factors that influence the house price most.

As a matter of fact, since the 20th century, companies have started using statistical methods to predict the future sales of their product in order to adjust sales strategy as well as optimize the unitality of the corporations. In recent years, with the help of machine learning thanks to the rapid development of computation ability, firms could predict the future sales of their products with a much higher precision based on the state-of-art scenarios and models. To be specific, machine learning models such as Decision Trees help firms predict the trend of their sales with high efficiency. On this basis and with this in mind, this study will investigate the implementation and application of advanced tools in sales prediction. With the appearance of neuron networks such as LSTM-RNN, firms are now able to predict with high accuracy. According to the analysis, this study concludes the result of research carried out on Decision trees, GBDTs, as well as RNNs. At the same time, the current limitations and prospects are demonstrated and illustrated.

Stock price forecasting has piqued the interest of academics studying finance and economics since it is a challenging but critical task for investors in financial markets. A variety of methods has been deployed, attempting to extract useful information to tackle the prediction problem. This paper presents rigorous research on the task of predicting the next-day price of Tesla stock by its past prices. A variety of methods are tested, including the long short-term memory (LSTM), the neural network model (NN), the autoregressive integrated moving average model (ARIMA), and the decomposition linear model (DL). The result is measured with mean absolute error (MAE). As shown in the experimentation result, the MAE of the ARIMA model is 7.2400, the MAE of the neural network is 5.8770, the MAE of the LSTM model is 6.8390 and the MAE of the decomposition model is 5.5890. The result suggests that the DL model performs the best in terms of MAE.

The development of natural language processing is significant for text emotion analysis because it helps to understand the expression of human emotions in different contexts and provides more accurate semantic understanding and emotion recognition capabilities for intelligent systems. In current natural language processing, sentiment analysis has become a key research field, and it is devoted to developing more accurate and efficient sentiment recognition models to adapt to the growing data scale and semantic complexity. This paper focuses on an overview of contemporary text emotion analysis technology and looks forward to the future development of natural language processing. This paper makes a detailed comparative analysis of the efficiency of different emotion analysis methods from the perspectives of key length, research content, research methods, and results. In the review, the advantages and limitations of various emotion analysis methods will be discussed in detail, including transformer-based and a series of the latest technologies. In addition, the performance differences of different methods of processing large-scale text data will be analyzed in-depth, and their performance in practical applications will be comprehensively evaluated. Finally, the research will discuss the possible future direction of natural language processing in emotion analysis in combination with current research trends and technology development trends to provide helpful enlightenment and guidance for researchers and practitioners in this field.