Universal artificial intelligence systems, such as OpenAI’s GPT, rely on large amounts of training data to improve the accuracy and performance of models. Research or medical AI applications often lack training data and computing power and can take advantage of new models designed to improve the efficiency, relevance and accuracy of AI output in more professional scenarios.
Large pre-trained language models (PLM) use increasingly larger data sets (such as Wikipedia) to train and optimize machine learning (ML) models to perform specific tasks. Although the accuracy and performance of large PLMs such as ChatGPT have improved over time, large PLMs cannot work properly when large datasets are not available or cannot be used due to computing limitations.
In short, a new artificial intelligence solution is needed to effectively leverage machine learning in research, medical or other applications where large amounts of information are not available to fully train existing artificial intelligence models.
To solve this problem, a team of computer scientists at the Singapore Science and Technology Research Agency (A*STAR) recently designed a collaborative knowledge injection method that can effectively train machine learning models using a small amount of training data. In this case, the researchers created a model that could more accurately determine the position or support or opposition of a particular target, such as a product or political candidate, based on the context of tweets, business reviews, or other linguistic data.
The team published their research on August 28 in the journal Big Data Mining and Analysis.
“Due to the diversity of targets and the limited availability of annotated data, position detection is essentially a low-resource task. Despite these challenges, position testing is critical to monitoring social media, conducting polls and informing governance strategies.” A*STAR scientist at the Center for Frontier Artificial Intelligence Research (CFAR) and the lead author of the paper. “Enhancing artificial intelligence-based low-resource gesture detection methods is critical to ensuring that these tools are effective and reliable in real-world applications.”
Smaller training data sets have a profound impact on the accuracy of AI prediction models. For example, the target “break the law” in Wikipedia is linked to a heavy metal song by the Reverend Judas, rather than the true definition of the term: acting in an illegal manner. Such erroneous training data can seriously affect the performance of machine learning models.
To improve the accuracy of AI gesture detection that relies on smaller training datasets, the research team focused on collaborative model mechanisms to: verify knowledge from different sources and learn selective features more effectively.
“Most artificial intelligence systems rely on pre-trained models developed using large predefined datasets, which can become obsolete, resulting in performance degradation. The approach we propose solves this challenge by integrating proven knowledge from multiple sources, ensuring that the model remains relevant and effective,”Ming said.
“Due to the large size of the parameters, pre-trained large language models also require a large amount of annotated data for training. Our method introduces a collaboration adapter that contains a minimum number of trainable parameters,… improves training efficiency and improves feature learning capabilities,”Ming said.
The team also uses hierarchical optimization algorithms to optimize the efficiency of large-scale PLM.
To test their model, the researchers conducted experiments on three publicly available gesture detection datasets: VAST, P-Stance, and COVID-19-Stance. The performance of the team model was then compared to the performance achieved by TAN, BERT, WS-BERT-Dual and other AI models.
Measured by F1 scores (ML model accuracy), the research team’s new gesture detection model for low-resource training data consistently scored higher than other AI models using all three data sets, with F1 scores ranging from 79.6% to 86.91%. Currently, an F1 score of 70% or higher is considered good.
The new attitude detection model greatly improves the practicality of artificial intelligence in more professional research environments and provides a template for further optimization in the future.
“Our main focus is efficient learning in resource-scarce real-world applications. Unlike major AI companies that focus on developing generic artificial intelligence (AGI) models, our goal is to create more efficient AI methods that benefit both the public and research, said Joey Tianyi Zhou, CFAR chief scientist and co-author of the paper.
Ivor W. from the Frontier Artificial Intelligence Research Center (CFAR) and the High Performance Computing Institute (IHPC) of the Science and Technology Research Agency (A*STAR) of Singapore. Tsang also contributed to the study.
For more content, you can click on the link below the video
Thank you for watching this video. If you like it, please subscribe and like it. thank
Original text:https://techxplore.com/news/2024-09-ai-stance-efficiency.html
Oil tubing: