As Novac/Ljubica Vuko writes, the Croatian startup airt, founded by Hajdi Cenan and Davor Runje, presented a new algorithm for learning and predicting behaviour from structured data, which is most often used in the business world, and applied for a patent for its global protection.
As they explained, the Croatian startip airt is building a platform for creating predictive models based on structured data such as, for example, that held by banks or communication service providers, and to process this data internally developed their own deep learning techniques inspired by methodologies used in language processing (NLP/Natural Language Processing).
They used their prior experience they had of working on specific problems from the financial sector to build a fully automated platform for the preparation of the transaction of the data and an automated model building for specific business problems.
In order to compare the quality of the platform, they decided to test and compare it with TabFormer, a system for the same purpose developed by IBM and for which it publicly published a synthetic data set for its testing. The initial test showed that the Airt model surpasses IBM’s (F1-score 0.90 vs. 0.86). However, although the accuracy of the model prediction is important, they emphasise that it isn’t the most important item for them.
”We believe we can do better than this result in terms of accuracy, but our primary focus isn’t on making the most accurate and precise model, but on reducing the resources needed to build one such model automatically. The greatest successes of deep learning techniques have been achieved in the fields of image and text processing, and to make only one such top model requires thousands or even tens of thousands of dollars for the electricity used to make them,” said Davor Runje.
He added that for such applications, it isn’t a big problem because one model is enough for each language, however, when it comes to (many) models used in business, it is clear that there are few companies that could afford something like that.
”Our goal is to achieve almost identical results as these expensive models, but for much less money in order to make our solution available to everyone, from the smallest web shop to the largest financial institutions,” said Runje.
With this approach, the Croatian startup airt has entered the “deep tech” domain, because the solution they’re developing is based on significant scientific and engineering challenges.
”We’re intensively engaged in research in and developing our own approaches and techniques. It was with the wholehearted help of Mladen Vukmir and William Zupancic from Vukmir & Associates that we submitted our first patent, which first goes to the EPO (European Patent Office), and then to the USA, for our own deep learning techniques on structured data. This is just the beginning because we aren’t going to stop innovating,” said Heidi Chenan.
The co-founders of “airt” will also say that with this approach and innovation they are trying to improve the side of deep learning that isn’t talked about too much yet, and that is the impact on the environment.
They say modern AI models consume an extremely large amount of energy. The computing resources needed to create the best models are increasing exponentially, doubling every 3.4 months, that is, in other words, in the period from 2012 to 2018, they increased as much as 300 thousand times.
”We’re aware of the trace that deep learning leaves on ecology and how, if this trend continues, this technology can become an opponent in the fight against climate change. Therefore, we’re working intensively to ensure that our system, in addition to scaling to the amount and speed, uses as little computer resources and energy as possible to process these large amounts of data,” explained Cenan.
It’s worth adding that digital transformation is one of the European Union’s top priorities.
For more, follow Made in Croatia.