How half an apple becomes a whole apple

Podcast with TOLERANT Software Managing Director Stefan Sedlacek

In a podcast interview, TOLERANT Software Managing Director Stefan Sedlacek revealed how fault-tolerant search works in data quality tools and the technology behind it. The podcast interview was conducted by Ashley Steele. You can listen to the podcast in German here.

Ashley Steele: TOLERANT Software – an established company with a clear vision. Address the right customers in the right way. Quality assurance of customer data is the core business of TOLERANT Software. Artificial intelligence is a hot topic at the moment and is frequently discussed. But what exactly is artificial intelligence? Which technologies are used? And what are the opportunities and risks? Managing Director Stefan Sedlacek gives an insight into this topic today and reports on how and since when TOLERANT Software has been using artificial intelligence to solve the problem of customer data quality.

Ashley Steele: Hello Stefan, how are you today?

Stefan Sedlacek: Hello Ashley. Thank you very much. I’m doing very well. Thank you very much.

Ashley Steele: Nice to meet you. We have chosen a very interesting topic today.

Stefan Sedlacek: Oh, yes, I think so too.

Ashley Steele: On the one hand, the topic of artificial intelligence in general, as I said in the intro. It’s very topical at the moment. It’s being discussed a lot. And then on the other hand, TOLERANT Match, customer data quality and the use of artificial intelligence in your products. So it’s a broad, exciting topic. But let’s start with the topic of artificial intelligence in general. What are the motivations behind artificial intelligence? What is actually happening at the moment and what is behind the current hype surrounding artificial intelligence?

Stefan Sedlacek: Yes, what motivates people in the context of AI? That’s a good question. I think the topic has only been on people’s minds for a few years. The development has been around for a long time. However, very few people were interested in it at the beginning because the developments were initially still taking place in research and academia. In a relatively short time, say three or four years, the applications have left the research field and become accessible to a wider audience. I would cite DeepL as an example. I think everyone knows this translation tool. It has already caused quite a stir that there is a translation engine that works much better than what Google has made available.

Ashley Steele: And ChatGPT is the other buzzword or application that is relatively new to the market and is being used by a wide audience.

Stefan Sedlacek: Exactly. In the last year or so, even more interesting applications have been added, such as ChatGPT, as you mentioned. A machine that can answer questions like a human. People are naturally curious about that. Or image generators that can create artistic images from descriptive text. You no longer need any special “skill” to create beautiful images. You can achieve really great results with just a few words and knowledge of AI.

Ashley Steele: It’s impressive what you see in this area.

Stefan Sedlacek: Yes, that is extremely impressive. And of course people are now asking themselves how far this can go. To what extent can I personally benefit from it? What risks does it pose for me and my environment?

Ashley Steele: Yes, I was just about to ask that. On the one hand, it’s about generating an image, for example a hill with flowers and a horse. And that is then generated. Or you give ChatGPT a text or a CV and it is then enhanced. These are things that are also beneficial for the average consumer (if I may put it that way). This is exactly the point you mentioned. What are the dangers of artificial intelligence? Will I lose my job? Will there be a change in the labor market? Or are there other dangers?

Stefan Sedlacek: I believe that the risks are still being evaluated at the moment. There are quite a few theories, especially on the subject of job losses. Some studies assume that the use of AI and robotics can automate many professions and that this will lead to high employment losses. This can be criticized for being mainly theoretical considerations. It does not necessarily mean that there will really be a loss of employment as a result of using a new tool. It can also mean that the new tool, i.e. AI, will enable more work to be done in less time. And more efficiency. How efficient AI is and how much work there is in volume is undetermined, which does not mean that there really has to be a finite amount of work. In this respect, as far as the question of job losses is concerned, the danger is perhaps there in certain areas, but it should not be overestimated.

Ashley Steele: There will perhaps be a change, but then the question also arises: If you move away from these dangers, what opportunities does artificial intelligence bring with it? What change will there be?

Stefan Sedlacek: Yes, the opportunities are of course to be seen behind the new fields of application that can be opened up with AI. The fact is that speech recognition and speech output would not work as well as they currently do without AI technology. We really do have a whole new field of application here. AI helps me personally as an artist or user with image generation, especially in the field of prototyping, for example. That didn’t exist until recently. AI is a great tool for producing results quickly and in a short space of time. Of course, you can also generate more sales if you can work faster.

Ashley Steele: And that also fits into this general topic of agility and prototyping and so on. Trying things out quickly. Getting customer feedback. That certainly helps in that area. But then perhaps a provocative question from my side. Or two questions. We know our way of life. Can AI become a way of life? How clever can AI become? Can AI be cleverer than us humans? Can AI think better? Is there this danger or not?

Stefan Sedlacek: That really is a very interesting question. If you actually read the headlines in the newspapers sometimes, there are also people who warn that AI could develop into a life form. At the moment, I would simply say that everything that goes in that direction and all considerations that go in that direction are to some extent marketing. Even statements from developers that AI can be worse than a nuclear bomb really have to be seen as marketing, in my opinion. People want to tell us that what they have created is so bombastic that everyone should really take a look at it. Based on current technology, I would say that an AI cannot evolve into a life form.

Ashley Steele: So I don’t need to be afraid at the moment.

Stefan Sedlacek: No, don’t worry. A typical tool like ChatGPT is always triggered from the outside until it produces a result. ChatGPT therefore works very deterministically and is not self-reflective. Of course, one could now consider whether such a language model should also be given self-reflective functions so that it questions, expands and improves itself. This has probably already been done to some extent, but even such functions would not breathe consciousness into a neural network like ChatGPT.

Ashley Steele: Let me interrupt you for a moment here. You mentioned a couple of buzzwords, like ChatGPT. You talked about language models. And you talked about neural networks. Let’s take three steps back. Is ChatGPT an application or a technology? And if ChatGPT is an application, what technologies are behind it? You’ve just mentioned a few buzzwords. Can you please explain a bit how language models and neural networks work from a technological point of view? Perhaps using an example of ChatGPT or another example.

Stefan Sedlacek: Okay, so ChatGPT is of course an application. It is not a technology in itself. ChatGPT is based on language models that are coupled with neural networks. So, now the question: What is a language model? A language model is – basically speaking – a computer program that understands natural questions and can also generate them itself. It is based on a statistical model that recognizes patterns in text and language data and then uses these patterns to evaluate the texts and predict future texts or language data. With this analysis from these language models, ChatGPT jumps into its pre-trained neural networks. Now the question: What is a neural network?

Ashley Steele: Exactly. You just read my mind, Stefan. Exactly. What are neural networks?

Stefan Sedlacek: This is also a term from IT, which is about being able to analyze large amounts of unstructured data well. Neural networks are actually capable of evaluating such large volumes and finding patterns in them. Typical unstructured data are images, videos, sounds, i.e. all the data that we produce in large quantities in our everyday lives. And the term “neural networks” is borrowed from biology. We have something similar in the human brain, i.e. a neuron, a nerve cell that is connected to other nerve cells and transmits signals to the other neurons or nerve cells depending on how they are connected. An artificial neural network works in a very similar way. Here there is a simulated neuron in the form of a mathematical formula that processes an input and generates output outside. Many artificial neurons work together to create an artificial neural network. You can say that a node in the neural network reacts to a value or to a stimulus that comes from outside. It then passes on a new value according to a certain calculation rule. Many such values and nodes then produce an overall result that ultimately allows us to decide whether an image shown is an apple or a pear.

Ashley Steele: Aha, okay, got it. You did a great job explaining that. I understood that there is more than just a language model and more than just a neural network. But it’s also about associative memory. That also plays a role in AI, if I’ve understood that correctly. What is associative memory? Is it better? Is it worse? Is it something else? Or are the three things combined together? Can you explain a bit how these three terms relate to each other?

Stefan Sedlacek: Exactly, in addition to the hype topics of “neural networks” and “language models”, there are other methods in the field of AI. One topic is associative memory. This works somewhat more simply than a neural network. An associative memory is a content-addressable memory. In other words, a form of memory that works with associations of content in order to access individual memory contents. Another way of describing it would be to say that a memory content is accessed by entering a memory value or a memory address. This is quite computer science-heavy, but let me explain what the advantage is. The advantage of this technology, as it is used in TOLERANT software, is that you can access results even if the input data is not complete. Even if – to stick with the apple example – you only had half an apple, the system would deduce that there was a whole apple behind it.

Ashley Steele: Okay, let’s take this example a little further using TOLERANT Match. If I’ve understood correctly, TOLERANT Match uses an associative memory.

Stefan Sedlacek: That’s right. Exactly. TOLERANT Match itself has associative memory as its technological basis. For us, this technology has a slightly different purpose than neural networks, for example. Each technology has its own strengths for its own area of application. The technology of associative memory – as we use it – fits very well with structured data.

Ashley Steele: Sorry, may I interrupt you for a moment? TOLERANT Match is about the quality of customer data, of address data? For example: Is my first name spelled A-S-H-L-E-Y or A-S-C-H-L-E-Y? The spellings are similar, but not the same. Thanks to the use of AI based on associative memory, such inconsistencies, i.e. matches that do not match 100 percent, are recognized and then actually found. Did I understand that correctly?

Stefan Sedlacek: You explained that very well. That is indeed the case. The basic problem is always that you want to find the right information for a query from a large pool of data. This is then a sub-question: “What is the correct address for a searched address?” Or: “Is a person identical to a large list of politicians?” Or: “What is the correct manufacturer’s name for a vehicle if someone enters ‘3 Series BMW 2-door 2007’ in an Internet form?” The associative memory technology then accesses large amounts of structured data in the background and can actually generate the correct answer from these very simple queries.

Ashley Steele: So that means the results are better with TL Match AI technology? Is the probability that the customer queries will be interpreted correctly even with typing errors higher with TL Match?

Stefan Sedlacek: That’s right.

Ashley Steele: But then a question: How does an AI system learn? How can you check whether the results of the AI system are good or bad? How does the learning process of an AI system work? What is the process behind it? Can you really be sure that the results provided by TOLERANT Match are correct?

Stefan Sedlacek: That’s a very good question. In conventional AI systems with neural networks, a relatively large number of training sections have to be inserted. To do this, you present the system with data whose type and structure is already known. In principle, the result is predefined. This allows the system to learn how to process data. If the system then spits out incorrect results, the processing function in the individual nodes is adjusted by means of appropriate feedback.
With TOLERANT Match, it’s a bit simpler. We only have one training step. We take the data that we receive in a structured form and break down the information we receive using a mathematical calculation formula so that it fits into our associative memory matrix. Although this is also a training step, it only needs to be carried out once. With neural networks, you have much longer training sections.
So how do you know whether the results are good or not? This is actually a project that is still at the research stage in computer science. You don’t know exactly because the many training sections in the AI systems change the nodes in such a way that the results are neither predictable nor comprehensible. So you have to test the system. You have to be confident that these tests are so accurate that you can ultimately say that the system is configured correctly.

Ashley Steele: So this is still an ongoing process. So you don’t just test once. Okay, you said at the beginning of the intro that the AI hype has only been around for 2-3 years. How long have you been using artificial intelligence in TOLERANT Match? Also only for the last 2-3 years? Did you do something else before that and then jumped on the AI hype? How did you do that?

Stefan Sedlacek: No, we’ve actually had this topic on our radar for a long time. We’ve been using it since the company was founded, in other words since 2009.

Ashley Steele: Since 2009? So you were pioneers in the use of these technologies here.

Stefan Sedlacek: Yes, so to speak. We had a predecessor company that did something similar. The topic was forced upon us. I first came into contact with the topic when the zip codes were changed in Germany. That was a long time ago. That was in 1992/1993, when there were already many address systems and the question was: “How do I go from a 4-digit to a 5-digit zip code?” That was indeed a tough change for many companies at the time. Of course, a lot has happened since then in terms of technology, but back then it was a real struggle. And that’s when this topic first came up. That was the first time I came into contact with it.

Ashley Steele: If I reflect a little, it’s a very complicated topic. You have – from my point of view – explained well what’s behind it. I hope the audience will see it the same way. What I found interesting at the end is that you didn’t wait until 2020 to use the technology, but have been using AI since the development of the TOLERANT Match and TOLERANT Move products, i.e. since 2009, as you said. You really are pioneers in that respect. I also believe – as you can see from your customers – that it was the right decision to rely on such technology and integrate it into the software.

Stefan Sedlacek: Yes, we are also very proud of that.

 Ashley Steele: Stefan, thank you very much. That was a very technical and complicated topic. But you explained it really well and clearly. I’ll leave it at that. See you in a few days and until the next TOLERANT topic we can talk about.

Stefan Sedlacek: Thank you, Ashley, I look forward to our next topic.

Ashley Steele: Thank you, Stefan, all the best.

Stefan Sedlacek: All the best, take care.