The TiRex model is smaller than the competitor models with only 35 million parameters and is therefore very memory-efficient. It is not only superior to other methods in terms of prediction quality, but also much faster than them.

“We are no longer talking about small improvements here, but about a significant increase in quality with TiRex compared to other models, both in the short and long term,” explains Hochreiter, Chief Scientist at NXAI and Chair at the JKU Linz.

Time series and their analysis determine everyday industrial life all over the world - including in robotics. "Many people look at large language models such as ChatGPT, but the great potential in industry lies elsewhere; for example in fast and memory-efficient time series models - in cars, in machines, on conveyor belts, during welding or in robotics. Everywhere there is time series data that can be used and with which money can be earned or from which digital products can be developed. Pre-trained time series models are downloaded millions of times today and are already being used commercially," explains Albert Ortig, CEO of NXAI.

The NXAI model TiRex relies on in-context learning. "This learning method allows zero-shot predictions - in other words, predictions for new time series require no additional training. This means that the model can also be used as a forecasting tool by non-experts and can be easily integrated into existing workflows. It also shows improvements in prediction quality, especially when data availability is limited," explains Andreas Auer, Researcher NXAI.

This allows machine builders, for example, to offer their customers TiRex models as digital products for optimization, which then run at the customer's site without any new training thanks to in-context learning and are automatically optimized based on the customer's data. “The decisive factor is how well the model generalizes to new time series - TiRex does this extremely well,” adds Hochreiter.

The model's ability to continuously monitor, analyze and update the system status is crucial for this. This so-called “state tracking” of the system status is not possible with transformer-based approaches. This means that TiRex can approximate hidden or latent states of a process over time, which improves predictions - for example, what is the robot arm doing? However, according to NXAI, the model architecture has another advantage, as it can be adapted to hardware and thus enables embedded AI applications.

The first roboticists are already testing the model and the xLSTM. We are awaiting the first use cases.