Small, fast AI models are shaking up business as we know it
In life, timing is crucial.
This is especially true in business, where every organization must forecast sales, demand, revenue and capacity requirements. Accurate and reliable time-varying forecasts can help any organization save (and earn) billions of dollars.
Time series forecasting is the foundation of what drives a business. It involves predicting future values based on past observations collected in constant time intervals, whether daily, monthly, quarterly or annually.
Artificial intelligence is expected to accelerate and sharpen business planning with new, faster and smaller base models designed for multivariable time series forecasting. These models don’t have to be the equivalent of an AI sledgehammer to get results. Small time series-based models or other small base models trained on high-quality, composite data are more energy efficient and can produce the same or even better results.
Vice President of AI and Automation at IBM Research.
How can time series AI models predict the future?
Time series models can be built from scratch or modified from existing pre-trained models and are best used for predicting outcomes in time series data. Traditionally, AI’s large language models calculate relationships between words to identify patterns in the data that can be projected forward to make better decisions.
Basic time series models look for patterns in historical observations to ‘understand’ a temporal process. These abstract representations ensure that the models can solve predictive tasks. The longer the time series, the better the prediction.
However, these types of measurements come with complications in a way that words, code and pixels do not. First, time series data is often continuous: think of video streaming from a self-driving car, temperature measurements from a reactor or heart rate data from a smartwatch. There is a lot of data to be processed, and its order and direction must be strictly maintained.
Time series data varies widely, from stock prices and satellite images to brain waves and light curves of distant stars. Compressing diverse observations into an abstract representation is a huge challenge.
Furthermore, different sets of time series data are often highly correlated. In the real world, complex events arise from multiple factors. For example, air temperature, pressure and humidity work strongly together to influence the weather. To predict a hurricane, you need to know how these variables have affected each other in the past to understand how the future might play out. Calculations and cross-channel correlations can quickly become overwhelming as the number of variables increases, especially when dealing with a long historical record.
The further back you go, the more complex these calculations become, especially if your target variable is influenced by other factors. For example, home heating sales may be related to erratic weather or the economy. The more interacting variables in a time series dataset, the more difficult it can be to isolate the signal that predicts the future.
Breaking barriers in time series forecasting
Basic AI models designed for time series forecasting can be difficult to build. The enormous scale and complexity of multi-channel data sources combined with external variables pose significant architectural challenges to the resulting model and non-trivial computational requirements, making it difficult to train and update models with reasonable accuracy and the desired prediction window in a timely manner. Today, many fundamental models cannot pick up the trends that rapidly evolving data patterns reveal – a process known as ‘temporal adaptation’. Basic time series models such as MOIRAI, TimesFM and Chronos are built on hundreds of millions of parameters that require significant computational resources and runtime.
The next wave of innovation
Researchers and practitioners are working on new ways to overcome these obstacles and unlock the full potential of using AI in time series forecasting. Can smaller models, pre-trained on only limited public, diverse time series datasets, yield better prediction accuracy? The answer turns out to be yes!
Experiments are now in full swing with the development of ‘small’ foundation models that are considerably smaller than 1B parameters. Smaller time series forecasting models (parameters from 1 million to 3 million) can provide significant computational efficiency while still achieving state-of-the-art results in zero/pair-shot forecasting, where models generate forecasts from unseen data sets . They can also support cross-channel and external variables – crucial features that existing popular methods lack.
These fast and small, general purpose, pre-trained AI models can be quickly deployed for use cases such as predicting electricity demand. They are also flexible enough to be extended to time series tasks other than just forecasting. For example, in anomaly detection, these small models can be trained on datasets containing anomalous and regular patterns, allowing them to learn the characteristics of anomalies and detect deviations from normal behavior.
We’re increasingly seeing that these small models, combined with business data, can have a big impact and deliver task-specific performance that rivals large models at a fraction of the cost. They are poised to become the ‘workhorses’ of enterprise AI.
In the coming years, AI is expected to help bring about a radical transformation in the business landscape. While most of the world’s public data powers current models, a vast majority of enterprise data remains unused. Small, fast foundation models – with flexibility, low development costs and broad applications – are poised to play an important role in this shift.
We have the best AI tools available today.
This article was produced as part of Ny BreakingPro’s Expert Insights channel, where we profile the best and brightest minds in today’s technology industry. The views expressed here are those of the author and are not necessarily those of Ny BreakingPro or Future plc. If you are interested in contributing, you can read more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro