AI is more than a zero-sum game

31.08.2018 – Wilhelm Kleiminger

 

Expert article for Computerworld dated 31 August 2018

Computers have long been better chess players. Nowadays they also steer cars and make clinical diagnoses. Is human intelligence about to be replaced by «artificial» machine intelligence?

Men have been fascinated by intelligent machines since long before the Austro-Hungarian inventor Wolfgang von Kempelen built a so-called chess robot in 1769. It was in fact a trick, as the device was controlled by a human player hidden inside. Interestingly, even though the computer pioneer John McCarthy created the concept of computer intelligence (AI) in 1955, there is still no clear definition that decides when we can talk about «artificial intelligence» when referring to computer programmes. The concept is often described in rather fuzzy terms as a successful imitation of human behaviour – using the so-called Turing Test. For this to apply, a machine must be able to sense its environment, think logically and interact with its surrounding.

Weak and strong AI

The fuzziness of the term AI is partly due to the fact that, for a long time, researchers could not decide whether artificial intelligence can be achieved «top-down» (starting from how the human brain works) or «bottom-up» (by creating expert systems). One common scientific classification distinguishes between two levels of artificial intelligence. «Strong AI» is the vision of a human-like intelligence, whereas «weak AI» is restricted to solving specific problems. For example, the «weak AI» definition covers Google‘s «AlphaGo» Go player and IBM‘s «Watson» semantic search engine, but also optimisation problems such as SBB timetable and track scheduling for better rail transport network loadings via traditional algorithms.

After initial success with formally describable problems like chess, it appeared that machines mainly have trouble with tasks that people find easy. In spite of some progress, machines still find it hard to understand spoken language. Even modern language assistants like Google Assistant need the person to adapt to the machine for communication. Unlike people, AI algorithms can compare millions of x-ray images in an instant, learning to identify the smallest tumours. This is a great help to doctors for diagnosing cancer.

The first artificial neural networks appeared in the early 1980s as a solution for this type of self-learning system – a solution based on the extreme simplification of the human brain by networking neurones as processing units. Machines can use this solution to learn to classify and universalise new information. Thanks to these new-fangled «deep-learning» solutions, Big-Data and – not least – the exponential increase in computing capacity, research has taken a big step forward in this area (often referred to as «machine learning») since 2006. The collective term «AI» now therefore covers solutions from the field of machine learning.

Potential for new applications

Solutions covered by the collective term AI are ideal tools for addressing a range of problems. An AI-based solution can help banks to analyse their payment transactions to detect money-laundering and other kinds of fraud. The system recognises and universalises fraud patterns from large databases and learns with every new instance. Interacting with networked objects, artificial intelligence can also generate added value for humans. In the field of building automation, heating, ventilation and air conditioning sensors can be evaluated and analysed on a continuous basis. This enables the heating to be controlled to individual residents’ needs. The analysis of energy consumption can also warn property managers of any damage to the insulation at an early stage.

Statistical methods have been used in the retail sector for decades now for stock and warehousing optimisation. This is now referred to as «data science» or «data analytics». Further developments in machine learning offer the added value of going into seasonality, item links and other demand influencing factors in even more detail. If the customer is interested, tailored services can be provided based on an analysis of their purchasing behaviour.

Success with fraud detection

Before introducing AI technology it is advisable to assess which problem actually needs solving and what data is available. Just like the heuristic model of Occam’s razor, simple AI solutions can be of initial benefit, as they are easier to validate. So when it came to finding a fraud detection solution for banks, analysing the available data using statistical methods was worthwhile at an initial stage. The findings were added to the AI system as initial training data so that it could develop the necessary instincts that would enable it to identify further instances on its own. The use of AI led to a higher detection rate with fewer false alarms («false positives»). Experts can improve the detection rate of today’s system with little effort at any time. They check complex cases manually and incorporate expert decisions in the training data. Man works closely with the system, achieving greater security for customers and bank alike.

Symbiotic relationship between human and artificial intelligence

At first glance, modern applications in the field of machine learning have little to do with von Kempelen’s so-called chess robot. But machines are already playing games of strategy such as chess and Go and quiz games like Jeopardy at the highest level. Learning systems detect whether a financial transaction is fraudulent and optimise retailer inventory. But we are still far from finding a strong artificial intelligence that addresses problems autonomously and implements the right solutions like people. Existing AI solutions – like the so-called chess robot – are ultimately complex overall systems and their «intelligence» is dependent upon the right combination of numerous individual components.

Human intelligence can be combined with its artificial counterpart to deliver real added value. People are not totally replaced with this form of collaboration. On the contrary: Man and machine support each other. But roles will shift in future. With the result that people – as has already happened with technical innovation in the past – can apply themselves to other more value-generating activities.