University of Georgia artificial intelligence predicts the future

An advanced model of machine learning designed by researchers at Georgia Tech has the potential to provide superior performance in a variety of areas, including disease protection, managing electricity consumption in cities, and boosting business growth. This new model, which is referred to as large pre-trained time series model (LPTM), as a single fundamental model can effectively perform forecasting tasks in various domains.

According to Tekna Technology News media innovation service, this model not only has the same or even better performance than existing dedicated models, but also requires 40% less data and its training time is 50% less than existing standards. In some cases, LPTM is even applied without the need for training data. Researchers have pre-trained the model on various datasets from various industries, including healthcare, transportation, and energy. To make effective use of this data, the Georgia Tech University group has developed an adaptive segmentation module.

LPTM will soon be presented at the NeurIPS 2024 conference in Vancouver, Canada, and its research draft has been published on the arXiv server. B. Aditya Prakash, one of the developers of the model, noted the importance of the work done in this area, saying that previous models were mainly trained on text and image data, but not much research has been done on time series data.

The underlying models, which are typically trained from different data, are capable of performing different tasks. These models, like the GPT and DALL-E engines, are popular generative AI platforms today. But LPTM is different in this regard, as it is designed for time series data, not text and image generation. Georgia Tech researchers have trained LPTM using data from epidemics, macroeconomics, energy consumption, traffic and transportation, stock markets, and human movement. After training, this model showed outstanding performance compared to 17 other models in various tests and performed best in five datasets.

In one of the experiments, the group of researchers tested the LPTM without the need for specific data and using only the input data, and it outperformed other models in all criteria. This outstanding performance shows the high potential of the LPTM model to predict superior results in various fields. Prakash, a professor at Georgia Tech University, pointed out that the LPTM model goes beyond forecasting and is capable of other tasks related to time series data, such as classification. He explained that traditional models are usually designed for each specific application because the data in each domain has different characteristics.

One of the salient features of LPTM is its adaptive segmentation module, which can handle temporal differences in data. This module is able to divide the data into different parts and choose the best methods to learn useful patterns.

The amazing performance of this model in tests has made it accepted in the NeurIPS 2024 conference. This conference is one of the three main events in the field of artificial intelligence and machine learning research. In addition to presenting their research at the conference, Prakash and Kamarthi, collaborators on the LPTM project, have also published an open source library of time series modules on GitHub.