The field of artificial intelligence with their manifold disciplines in the field of perception, learning, the logic and speech processing has made significant progress in their application over the last ten years. The decisive factors for this progress were: the quantity of available data, the increase in computing power, the availability of free software development environments and the further development of new algorithms and architectures from the environment of machine learning. Systems from the field of artificial intelligence and its algorithmic decision patterns influence more and more elements of everyday life: The interaction in the household through home assistants, the relevance of search and advertising offers, re-shaping mobile driving and traffic management, the diagnosis in medicine or also the allocation of the personal credit line.
“It’s not a human move. I’ve never seen a human play this move.”(Fan Hui on the 37th move of the Go game between Lee Sedol and AlphaGo, 2016)
The outstanding successes, especially in the area of supervised machine learning and here especially the aspects of representative learning and deep neural networks, so-called deep learning, have significant impact not only in the academic world, as in the outstanding milestone in the Go game between Lee Sedol and AlphaGo, but also show their added value in industrial applications – Industrial AI. Algorithms make our lives more efficient, increasingly support our decision-making processes, and sometimes even take over them. From speech recognition and translation, visual quality inspection, dynamic pricing, autonomous parameter optimization for data centers, improved adaptivity of robotic controls and AI-optimized supply chains to predictive maintenance in production. We are now in the area of putting Industrial AI in operation – towards an focused and pervasive use of AI & ML across application fields. The impact of Industrial AI thrives with the human factor as main discriminator following the leading principles of autonomy and assistance, by connecting the virtual and the physical world, which drives its value to proposition leading to new insights at reduced costs for our partners and customer.
In contrast to the industrial environment AI (B2B), which has always focused on increasing efficiency and productivity, the consumer sector (B2C) focuses primarily on aspects of predicting behavioral patterns and optimizing the attention horizon of the customers. From the placement of advertisements to the automated filtering of news articles up to the AI-supported image review. The strength – and at the same time the danger – of using machine learning methods such as deep learning to predict or classify new situations is, that the quality of the result depends heavily on the size, balance, and purity of the input data.
The transferability of the model to a new problem is therefore very difficult to assess. While in statistics the concept of significance is used to assess the significance of each analysis in each of its steps, which are critically questioned, AI lacks such a punctured measure of quality for an algorithm. As a rule, validation can only be performed based on the quality of the result on the corresponding training and/o test data. However, whether correlations – within these multi-layered neural networks and despite its large data – occur by random (perhaps because errors cancel each other out or the applications do not expose the relevant distortions) or are grounded on meaningful pattern stays often hidden. Accordingly, such algorithms lead to, due to wrong assumptions in the learning process, such as a low diversity of data sources, faulty robustness in changing application domains or wrong assumptions in the modelling process, unbalanced results such as in the case of gender-discriminatory allocation of credit lines.
“One of the first things taught: […] correlation is not causation. It is also one of the first things forgotten.”(Thomas Sowell, Stanford, 1930)
It is the nested non-linear structure of modern deep learning systems, which makes it difficult for users and AI experts to gain transparency, which information or features have been used by the system to come to its decision. It is difficult to assess, which information may only a random correlation, but may not allow for significant causality. Therefore, these types of AI system are often referred to as „black box system”. Impairing faulty results then directly the profitability. However, as soon as decisions are made based on the recommendations of an AI system that have a significant, perhaps even existential impact on individuals or groups of people, the requirements for such a non-discriminatory algorithm are naturally much higher and the error tolerance much lower. So, how can we ensure that a system can also be used in a real environment functions as expected and meets the high requirements for non-discriminatory results? How can the risk of getting bad, unfair and unstable results from a neural „black box” be reduced?
“Professional responsibility […] is not to discover the laws of the universe, but act responsibly in the world by transforming existing situations into more preferred ones.”(Herb Simon, 1996)
In a first approximation, of course, the currently rapidly evolving regulations play a major role in limiting the risks that can arise in the design and use of AI. In addition to existing product liability laws, the DSG VO or security regulations, however, many institutions and companies concretize their approach in charters or sets of rules that are intended to do justice to the special features of AI. At Siemens, we use a set of seven „mitigation principles“ (see figure below), which we believe help to harness the undisputed advantages of AI within a responsible framework.
In addition to sensible rules tailored to AI the holistic inclusion of different perspectives plays a major role as the second major lever in risk minimization.
For sure, we live in a volatile world marked by uncertainty, in which we all have blind spots in our perception and power of judgement and thus often decide in unconscious bias, makes it necessary to recognize and correct this human weakness from different perspectives, also in the field of artificial intelligence.
Diversity as an elementary building block in the AI life cycle: The data used to train AI systems must cover the range of their later use cases in order to arrive at valid results.
But different perspectives must also be considered in research, the development, and the applications of AI. From end users to domain experts to software development – we must value diversity in all its dimensions such as gender, social and ethnic origin as a social potential and integrate it. Diversity is therefore not only a motor for outstanding innovation performance, but also elementary for the reduction of bias also bias in artificial intelligence.
The implementation takes place in training, awareness, and engagement initiatives, short-term innovation formats such as hackathons, boot camps or innovation sprints, but also in dedicated locations for co-creation, which are intended to fulfil the claim of being a platform for different perspectives.
“Diversity is not only the motor for outstanding innovation performance, but also fundamental to reducing bias also bias in AI”.
An elegant and especially in the European context very relevant way to resolve the contradictions between the potential and the danger of AI are new technologies that allow a holistic view on the influencing factors of AI applications in their life cycles: data generation and selection, algorithm selection and explainability, accuracy and runtime, but also provision, updating and monitoring of its applications. There are already many relevant technological components available that can help to eliminate implicit prejudices in the data and thus also support to uncover AI-based prejudices and make them correctable. Furthermore, they contribute to meet the high requirements for the protection of personal data, without being hindered by the higher complexity in processes and products, and ultimately improve the robustness of the AI systems, which, concurrently, minimizes their susceptibility to errors and facilitates economically attractive scalability to other application fields. Currently, relevant technologies are:
- Explainable AI is a field that addresses the interpretability of black box decisions in AI. Explainable AI is defined in terms of the explainability of (for example, data, input characteristics), during (e.g. model architectures, relevance characteristics) and after (e.g. test and target references) of the modeling. In industry, these methods are used with black-box approaches, among others, to accuracy of the AI algorithms can be explained. This helps to make the process more understandable for customers but also supports to illustrate internal system distortions.
- Active Learning is another emerging field in the AI, which not only describes the process of the AI with can speed up little labelled data, but the „human-in-the-loop-paradigm” within that shapes AI. Permitted in industrial applications this approach to integrate feedback from domain experts into the AI training cycle, i.e. to train the system through human usage behavior and domain knowledge continuously and efficiently to improve.
- Trustworthy AI aims at the area of trustworthiness and robustness of algorithms, which provide the user with feedback on errors, robustness, and or inconsistencies in all phases of the AI. Target is to enable AI applications to detect a possible domain change and the corresponding adaptive uncertainty and report back.
- Federated Learning is a distributed approach of machine learning, which is based on the model training large amounts of data from distributed edge devices. The basic idea behind it is to use the code to the data instead of sending the data to the code, and addresses here the basic aspects the privacy, ownership, and location of the data.
- Differential Privacy is a mathematical method for anonymizing records of the use of metadata, thereby ensuring privacy of the individual can be safeguarded. An algorithm analyzes a data set and calculates statistics about it (e.g. the mean value, the variance, or the median). It is called differential private, if you can’t tell from the output, whether the data of a person in the original record were included or not.
- Edge AI not only allows real-time processing of data collected on a hardware device („Edge“), but also allows the AI to act as a trustee, since data generation and analysis is performed within the Edge. This means that data can be processed, and decisions can be made decentral without the need for a connection to a central system (e.g. cloud).
However, a responsible use of artificial intelligence requires not only technical and institutional control and monitoring of the AI process regarding bias, fairness, transparency, accountability and explainability, but also continuous further training of developers, users, and decision makers. These must be able to understand the advantages, dangers, and procedures for risk reduction of AI methods and applications in Learn to know and evaluate trainings.
“Trust is not necessarily about transparency but about interaction. “
Trust and safety are the most important imperatives for people, process, and products throughout the entire life cycle of AI. Industrial AI targets industrial-grade trustworthy AI solutions that thrives for robustness, explainability and security. The benefits and consequences of AI are still unfolding and will continue to fundamentally change society and the economy. It is therefore even more important to shape technological and cultural change in a joint and responsible manner.
This article was first published at Handelsblatt Journal: “Ulli Waltinger and Benno Blumoser: Responsible AI: Transparenz, Bias, und Verantwortung in der KI, Handelsblatt Journal Future IT 2019”