Deep learning and AI are all the news today. Are those left behind who don’t jump on the bandwagon? Far from it, we hear from experts at Smals. In fact, deep learning is no better than older techniques in many cases.

Hardly a day goes by without news from ChatGPT. While more and more European countries have serious questions about the legality of the latest AI models, technology giants like Google and Microsoft want to integrate GPT-4 and Co. into their products. AI seems to be the solution to almost every business question. Or not? Vandy Berten, ICT researcher at Smals, offers some nuances.

future in simplicity

“ChatGPT will not be widely used in the coming years,” he predicts. “This is not necessary either. Relatively simple techniques such as classic analytics still have a lot of future.” Berten notes that many classic analytics projects still have to be rolled out, especially in administration. “We’re working on a lot right now network analysis‘ he hints. Berten does not refer to IT networks but to the analysis of relationships between different entities. This results in insights that can lead to optimization.

In practice, deep learning is no better than older techniques in many cases.

Vandy Berten, ICT researcher Smals

This choice of analyzes is not a limitation, he emphasizes. “If there is a problem, you have to look for a technology that can solve that problem. We don’t work the other way around. In practice, deep learning is in many cases no better than older techniques.” Katy Fokou from Smals AI research department agrees: “Classic analytics to monitor or understand phenomena is still very important.”

black box

In fact, deep learning brings with it problems that do not arise when examining data using classical analysis techniques. Specifically, it’s about the black box problem: What models like GPT-3 and GPT-4 do cannot really be explained. Researchers know the model works, but they can’t use ChatGPT’s thought pattern, for example, to explain why the AI ​​responded a certain way to a certain question.

This isn’t just a problem with Smals. Berten: “If an algorithm suggests something, we have to be able to justify this suggestion in concrete terms. We cannot rely on a black box and that is very difficult with current techniques. With classic analytics, you can easily indicate which parameters and relationships are responsible for specific insights.”

Explain delicate decisions

It will not surprise you and maybe even reassure you that under these circumstances it is very difficult for a party like Smals to just unleash AI on data. Smals provides the IT for the social security institutions, so the data is often sensitive personal data. It is inconceivable that they are processed by only one large language model (LLM) like ChatGPT in the cloud. “Our data protection officers are very careful,” emphasizes Fokou. “Accessing personal data takes a long time.”

Professor Isabelle Boydens, data quality expert at Université libre de Bruxelles and Smals, also points out that the black box problem and data sensitivity are not problems unique to Smals. “Even an insurance company cannot deny rights without reason,” she explains. The EU is currently working on AI legislation that will limit the use of inexplicable AI in many cases. Fokou: “It will be much more difficult to use deep learning with these new rules.” Clear, explainable and reliable findings from classic analytics are retained.

Not always a disaster

That doesn’t mean that Smals, or other organizations managing sensitive data, should look on with regret while the rest of the world embraces ChatGPT and the like with gusto. Play two things. On the one hand, the black box problem is not unsolvable. “Research within AI is moving towards algorithms that are less affected by this black box phenomenon,” says Fokou. “People want to know the reason for a certain result.”

Research within AI is moving towards algorithms that are less affected by this black box phenomenon

Katy Fokou, AI Research Department Smals

On the other hand, not all data is very sensitive and the black box is not always a big problem. “If algorithms are used in e-commerce to recommend things and they work well, it’s not a big disaster if the underlying AI’s reasoning is a bit unclear,” explains Fokou. “LLMs can also be interesting, for example to build up an FAQ.”

Unless, of course, the FAQ is based on sensitive data. “Right now, AI based on LLM’s US technology is in the cloud and is by no means transparent.” Boydens agrees. Within the EU, an AI chatbot is useful to help answer questions in all languages. There is certainly an advantage in writing FAQs or texts that have nothing to do with confidential information.”

Good framework conditions, good solutions

In the long term, Fokou believes that technology itself is not the problem. As with analytics, AI and deep learning need a clear framework. That is currently missing. In addition, many organizations, including Smals, are certainly not yet able to see the reason behind the analyses. Currently, LLM-based AI is a promising technology with many downsides, including transparency.

As Berten says, analytics is much clearer and easier to explain. Many classic insights can usually be gained from analytics, for example by linking new data sets. Such projects are in themselves a major challenge, which in the case of Smals is carried out with the necessary legal due diligence to protect your data from careless misuse. And then we haven’t even mentioned the importance of qualitative data, both for analytics and AI (or rather in this article).

Thanks to ChatGPT, AI is hip, but new, difficult to explain, not necessarily compliant with data protection and not mature anyway. Also, AI cannot see into the future. Analytics sometimes feels like artificial intelligence, but it is a much more mature technology that is currently producing better results. After all, as a business, you want to respond to insights from data you understand, not prompts from a third-party AI that you must blindly trust.