Leveraging AI Technologies for Porsche’s Future
Tobias Grosse-Puppendahl and Jan Feiling, both Porsche Innovation Management, talk about how Porsche balances the groundbreaking innovations through data collection for AI purposes and protecting the privacy of their customers.
Next to zero emission mobility, Artificial Intelligence (AI) is one of the most promising fields for the automotive industry and especially for its customers: Just imagine arriving at a meeting and having your car park and charge autonomously, without the hassle of even finding a parking spot. On your weekend, you could be provided with a joyful driving experience comprising a sporty drive along your favourite back roads and the optimal drive setup for your individual car.
These engaging customer experiences would not be possible without the recent mind-boggling developments in the area of AI. Today, anything from autonomous driving to intelligent traffic management systems or shared smart mobility is enabled by the use of AI-related technologies and methods, for example Machine Learning.
Over the last decades, the automotive industry has mastered the art of model-based function development, which is grounded on physical modeling with relatively few tunable parameters. With the shift towards machine-learning technologies which incorporate the ability to learn from experience, models approach human-level perception and actuation capabilities. However, they also tend to become more of a black-box with thousands of parameters that are trained by vasts amount of previously recorded data. This fact, and the growing customer demand for personalization, is one of our drivers to build new forms of machine learning techniques and improve upon existing ones.
Improving Personalization and Data Security by Design
In the past years, we could observe that data privacy and security is advancing to a luxury good: Possibly everyone of us trades very personal data against the offering of highly personalized services. At Porsche however, we want to improve privacy and data security by design. The big challenge we are tackling is: What if everybody could still benefit from the knowledge of the masses, without the need to reveal too much information about ourselves?
An example could be, the usage of personalized speech and haptic data to teach and improve a personal assist in our cars. Conceptually, when the personal assist shows a suggestion query, the car locally stores information about the current context and whether the customer clicked or accepted the suggestion. This data can be used to apply machine learning techniques to improve the recommendations of the personal assist. But how can every customer benefits from such knowledge, without violating the data privacy of our customers?
To solve this problem, we are actively investigating use-cases that employ an emerging technique in Machine Learning called Federated Learning [1]. With this technology, we can maximize data security and privacy, with the additional upside of significantly reducing data transfer between our cars and the backend infrastructure. The technique enables us to still contribute to a global set of knowledge for the benefit of other Porsche customers.
In contrast to standard machine learning techniques, Federated Learning represents a fully decentralized learning approach, with training being carried out on many distributed machines, in our case our customers’ cars. The magic sauce of Federated Learning is a method for merging thousands of local models into a global model by a negotiation protocol. This form of learning helps us to solve issues that are related to data which is massively scattered among a lot of individual clients as well as issues that are caused by data not being uniformly distributed. For example, there will be significantly more data available on wonderful Porsche-routes like the scenic Highway 1, compared to less frequently visited places.
Between Personalization and Swarm Intelligence
Today and in the future, we want to offer our customers a unique and personalized Porsche experience based on their personal preferences. It is part of our principles that data collection should never influence a user‘s privacy negatively. To achieve this, we have outlined how we aim to solve the contradiction between personalization, swarm intelligence and privacy by the use of Federated Learning. However, there is an important drawback because some use-cases will still require us to store bits of personal information in the cloud, instead of just within the car.
For example, revisiting the personal assist use case from before. Every customer helps by their locally stored and protected data to improve the so-called predictive models to improve their suggestion accuracy. Nevertheless, each client interacts differently with the system and there will be probably some suggestion queries which suits not everyones preferences. Considering the case when a customer says to the system, “I feel cold” — is it right to heat up by one, two or three degrees celcius? Who knows? This depends on the individuum, hence, on top of the shared knowledge by Federated Learning some personalization character is required.
Therefore, we are exploring new developments, one the so called multi-task Federated Learning, which has the advantage that each customer can access the collaboratively trained models, but influence part of this knowledge by their personal preferences locally. The main and important differences is basically how the model is trained and inferred in the car and how those locally personalized models are merged in the cloud. Each data source can be seen as one task, but there could be similar tasks, hence clients with same preferences. Thinking of the “I feel cold” use case from before — there will be many people which are happy with heating up by 2 degree celsius. Technically speaking, a wrapper model, which is stored in the cloud and accessible, like the collaborative models, by each customer, admits to identify relations between the locally stored data, without exchanging those. This wrapper model can then be used to parametrize the suggestion queries by each customer in a personalized way, since the output of this wrapper model depends on the locally stored data. Summarizing, multi-task Federated Learning is based on a collaborative trained knowledge source where each customer pick parts of a model by their own preferences automatically.
In regard of this novel and promising tool kit one can also think of a lot of use cases in the field of automated driving. A few examples are corner case detection in driving maneuvers or video and image segmentation, where large data streams are needed to train machine learning models on the new data. On one hand, Federated Learning shares this knowledge by its distributed character, on the other hand data transfer is reduced, since models are much smaller compared to data.
Safety must be our Primary Focus
When critics warn about artificial intelligence, they often claim that deep learning-based AI has lost a very important property of natural intelligence: the ability to explain its conclusions and decisions. In its efforts to advance autonomous driving, the car industry traditionally uses a model-driven function development based on physical models. These are combined with a set of a few parameters optimized mathematically or in experiments with extremely skilled test drivers.
But with the emerging deep learning techniques which provide tremendous cognitive abilities, such networks containing thousands of parameters are often regarded as black boxes. And very often, these black-boxes work completely different than our own visual cortex. For example, you may have heard about methods that trick deep neural networks into misclassifying a stop sign as a speed 45 sign by adding a few stickers, which is shown in Figure 1 [2].
Therefore, AI research is always directed by our ethical responsibility which guides the choice of our AI methods and architectures. To ensure the safety of our customers in each and every situation, we are carrying out research that help us to better understand complex neural networks. The first step towards understanding such networks is to understand the uncertainty of its output. For example, our colleagues at Volkswagen Group Research are actively investigating methods that deliberately drop random neurons within a network to estimate the statistical uncertainty of a network’s conclusions [3]. Another complementing approach our colleagues at Porsche are actively pursuing is to learn to interpret visual information much more like children do by the use of simplified visual representations [4]. Additionally, general advances in optimization theory, like algorithms which can be tailored for training machine learning algorithms and neural networks [5], developed by one of the authors of this article, is also a key enabler in extending state-of-the-art artificial intelligence research, applicable to more complex use cases in the future.
We also believe that blockchain and other methods for transparent, forgery-proof recording and tracking of transactions of all kinds could contribute here. At the same time, we can use blockchain to validate the trust of our customers with regards to their privacy, for example by storing the topologies of our neural networks in a public blockchain. Together with XAIN, the winner of our Blockchain Open Innovation Contest, Porsche was the first car manufacturer to bring the blockchain into the car.
An unprecedented amount of AI technology
In short, we consider AI as an enabling technology that spans horizontally across the whole organization — very similar to the impact that programming languages had on our work. For our customers, our ethical approach to AI provides the foundation for many desired functions like automated driving and highly personalized services with high levels of privacy.
While autonomous driving receives most of the attention currently, it is only one of many technologies. Personal Mobility Assistants are a typical field of application, including voice control in the vehicle, which has moved away from simple voice commands to natural-language dialogue capabilities — a feature that we have recently introduced in our new Cayenne. The upcoming, all-electric Taycan will present an unprecedented amount of AI technology — while at the same time making sure that our customers are fully protected in terms of privacy, safety and transparency.
Ferdinand Porsche lay the foundations for our culture when he said: “I could not find the sports car of my dreams, so I built it myself”. Fascination for sports cars, combined with the highest degree of engineering has led Porsche over the last decades. Our company and customer culture imply that there is no blueprint for us. With regards to our technological future, this means: „We could not find the AI of our dreams, so we built it ourselves“.
References
[1] Konečný, Jakub, et al. “Federated learning: Strategies for improving communication efficiency.” arXiv preprint arXiv:1610.05492 (2016).
[2] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaiowei Xiao, Atul Prakash, Tadayoshi Kohno, Dawn Song: Robust Physical-World Attacks on Deep Learning Visual Classification, Conference on Computer Vision and Pattern Recognition (CVPR). 2018.
[3] Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a Bayesian approximation: representing model uncertainty in deep learning. In Proceedings of the 33rd International Conference on International Conference on Machine Learning — Volume 48 (ICML’16), Maria Florina Balcan and Kilian Q. Weinberger (Eds.), Vol. 48. JMLR.org 1050–1059.
[4] Biehl, M, Barbara Eva Hammer and Thomas Villmann. “Prototype-based models in machine learning.” Wiley interdisciplinary reviews. Cognitive science 7 2 (2016): 92–111.
[5] J. Feiling, A. Zeller and C. Ebenbauer, “Derivative-Free Optimization Algorithms Based on Non-Commutative Maps,” in IEEE Control Systems Letters, vol. 2, no. 4, pp. 743–748, Oct. 2018.
