Comment by Benjamin Uncover, Datanomic
AI as a Service – How a company sells data without letting go
By Benjamin Uncover
providers about it
Data monetization not working? Or is it immoral? Quite the opposite, because in fact companies can sell the data generated in the company or even just flow through the company and thus advance the community.
Anyone who hears nothing but “sale statements” might first think of the tax CDs that the German state has already acquired several times, sometimes under questionable circumstances, and most recently again in 2021 from Dubai. Or perhaps the thought comes first to the many free social media platforms that, according to their terms and conditions, do not hide who owns the uploaded data from the moment of use, at least to patient readers. Nothing is free, just the means of payment does not have to be money, because the data also has value that can be assessed.
However, we do not want to go into further detail here about this wisdom that has already reached the awareness of the general public. Handling personal information should not become part of our culture. However, there is a legal and socially beneficial way for the company to turn data into valuable assets.
Selling data without being able to give it away
In particular, a company’s internal data is valuable because only it provides detailed information about products and processes. This includes in particular transaction or movement data, often saturated with personal references and trade secrets. As a rule, this data cannot be offered for sale, as it should not fall into the wrong hands – and often this data is more than direct competitors. And when it comes to people’s personal data, this is in any case impossible without the direct consent of everyone involved. So selling data – and thus also monetizing data values - seems impossible.
Turning data into gold with AI as a Service
Data that cannot be sold may still be made available to third parties. However, it is not the data itself that is then sold, but the generalization knowledge gained from that data—from the quantity and variety of individual datasets.
Prediction models feed from supervised machine learning from data histories in order to derive specific predictions from them. These prediction models can be generated using classic machine learning or deep learning methods. Deep learning includes artificial neural networks, which are a subset of machine learningboth are part of artificial intelligence (Amnesty International) or. Artificial intelligence (AI). AI models mature from data via what are called training-learning connections.
These pre-trained models can then be offered for use for a service fee (AI-as-a-Service) or, in special cases, and if certain requirements are met, sold out.
Users of the pre-trained models can then either use them directly or retrain them using their own data, thus adapting them to their own, but similar, application scenarios. when using Deep learning It is also possible to adapt pre-trained AI models into the architecture of retinal neurons in order to further align them with their new target. The concept behind it is called transfer learning.
Knowledge transfer is a boost to the economy and society
The data collected by companies represents operational experience about operations and their results. Transfer Learning AI feeds on this experience and passes it on to other companies in the same or different industries. Companies that have not had access to such expertise for many years – if at all – can now purchase it at a price in line with the market and thus expand their operational capabilities. Today’s data makes it possible to take shortcuts in creating experience and cheat somewhat in the game in the market.
Not only does AI-as-a-Service serve economic interests, but we also benefit from AI trained in medicine, which can draw effective conclusions for individualized treatment of patients from a large number of treatment and drug data.
Current areas of application of AI as a service and learning transfer
There are already countless application areas for knowledge transfer through AI as a service in industrial production as well as in retail, logistics, real estate, insurance and financial services.
The industrial trends of AI-as-a-Service can be found above all in the machine tool industry and automation technology. Here, machine system providers, for example, provide prediction models provided for energy consumption, output quality, and optimal maintenance times. In particular, the last use case, too Preventive maintenance Known, is an important additional work in machine tool construction.
In retail, AI-as-a-Service is already being used to improve purchasing and working capital through sales and warehouse forecasts, as well as revenue forecasting and better management of marketing campaigns (without triggering returns). Dynamic pricing models are also used.
In finance and insurance, pre-trained models are used to extract relevant data from reports submitted by applicants, brokers, and expert opinions, with huge efficiency gains. Insurance requirements for buildings or damage to vehicles can already be determined using AI through image recognition. Insurance companies and partner companies train these models and provide them to other companies.
Warning: Prevent data leakage!
Transfer learning is a powerful tool for companies to reuse AI models, but where there is light, there is shadow. Prediction models, which are sometimes trained on sensitive data, must be verifiably secured in a way that does not allow any conclusions to be drawn about individual datasets. It is not easy to reconstruct the original data from the trained models – at least if the models are already generalized. However, hacker attacks on deep learning models aimed at re-engineering sensitive data are to be expected in the future, and the topic is less extensively researched than some machine learning engineers are willing to admit. However, with experience and knowledge after the introductory course in deep learning, the risks can be reduced to an acceptable minimum.
In general, trained models should not be delivered directly to the buyer, but should be self-hosted and hidden behind the API wall, so that the buyer can only query the model via the API, and then AI as a service.
Article files and article links