Artificial Intelligence

Artificial intelligence plays an essential role in new scientific research, and it is expected that many AI-based applications will be available in the market in the future. Powerful hardware such as GPUs can perform complex statistical calculations and provide an acceptable solution for customers as a product.
A few years ago, we could not have imagined that a voice assistant like Alexa could be a product. But it is now integrated into many products, and there are many Alexa-compatible smart home devices, and Alexa compatibility is now an important feature.
At the end of the day, customers will decide whether AI-based products are mature enough and usable or still in the research and development phase. Making a product with an Ai Model is complicated because of the inherent uncertainties in AI models or are sensed as model inputs from the outside.
Model view versus system view
Most AI research is now done in the laboratories of universities or companies or sometimes in collaboration. Building a system based on AI research is the responsibility of the manufacturer, not universities or research institutes. About 90% of the AI models developed in laboratories do not enter the production line and are not used as part of a system as a product.
The current product lifecycle management of the systems does not include and address AI-based applications. We have completely separate lifecycle management for how the AI model is developed in the ECU software. As long as these two product life cycles run parallel, it won’t be easy to implement an Artificial Intelligence model that meets all the system’s hardware and software requirements.

Stability of an AI model and generalization
An application based on artificial intelligence with 95% accuracy in the laboratory is not stable enough due to the statistical nature of artificial intelligence models and the quality of training data whether all relevant scenarios for the application are considered or not.
AI applications must be able to detect out-of-distribution data that is not part of the training data. Giraffe filtering in an image processing system trained only to detect dogs and cats is an excellent example of out-of-distribution.
Generalization in Machine Learning means that the AI model adapts well to previously unseen data from the same distribution used to create the model. Previously unseen data can be very close to the classification decision line between groups, and sometimes you need to update the AI model to avoid misalignment.
Mission-critical problems
The implementation of systems with high complexity and uncertainty is not possible with rule-based software. But how can we transfer or redefine our development process in rule-based software to achieve safe and reliable AI-based applications?
Products based on artificial intelligence have an uncertain part that originates from their statistical and probabilistic background. Dealing with mission-critical issues is vital and difficult to achieve high confidence levels as it is currently used as "state of the art" in rules-based systems. We apply five Safety integrity levels from QM to D in the automotive industry, and this definition needs to be translated into an AI-based system and redefined.
An AI-based application like a spam detector is not a safety-related application and could be optimized to send more emails to the spam folder and is not critical. But if an AI-based pedestrian detection algorithm fails or is not accurate enough, causing a high rate of type one or two errors (false positives and false negatives, respectively), it is not mature enough to be sold as a product in the market.
Summing up,
AI system developers, mostly data scientists, should change their view from an AI model view to an overall system view. This means more attention at the system level, including hardware and software and their interactions, and considering the model’s value from the customer’s point of view for the overall system.
Domain knowledge is the most important topic for data scientists to develop a high-performance and safe AI model. We see the same situation with cloud computing applications in various industries. To have successful cloud solutions in the automotive industry, you need very in-depth automotive knowledge from the customer’s point of view down to the technical aspects.
The production of an AI-based system includes many components such as data collection, feature extraction, monitoring, verification, etc. Addressing all components is essential for the product to be successful on the market. Many decisions that are made in the very early stages are important for success in the end. For example, when deciding to use static or dynamic training, we should consider the tradeoff between the two and consider the pros and cons of both and which are more appropriate for our application.
Mission-critical problems that prevent a reliable and safe product with added value for our customers should be identified very early in the project development phase. At least one solution must be available and implemented in parallel to increase confidence in AI predictions and decisions to the acceptable level to society.
There is no doubt that we will have more and more AI-based products on the market in the future, and this market is growing very quickly. Performance, safety, and availability are the most important properties that most AI applications on the market should consider. To build a bridge between research activities and production, we need more cooperation in the industry and a redefinition of the entire development process and exchange of ideas.