Artificial intelligence (AI) is the simulation of human cognitive functions. Modern AI can do many tasks without human help, such as playing a board game, booking a trip, finding a venue, or writing an article. We have been working on AI for more than 20 years and helped shape the history, establish state-of-the-art approaches, and aim to solve future challenges of this fascinating topic.

Mid-20th century visions of artificial intelligence described a general AI that would simulate human intelligence in every aspect. This included the physical manifestation of humanoid robots, a fantasy that lives on in our attitudes towards AI. But early research strongly suggested that the development of general AI was a remote and unlikely goal. The rule-based methods used at the time were too complex to be used in real-world tasks. Comparable to using a recipe for cooking, real life application shows details and challenges that the recipe never dealt with.

Inspired by the way the human brain works, neural networks were invented. These are computer systems made up of many simple decision-making units that can be trained to do a task using examples. When showing such a network a large set of X-ray images with their respective diagnoses, it learns the relationship between the image and the diagnosis. Based on learned information and patterns, it can correctly diagnose based on new images. Nonetheless, X-ray images in diagnostically meaningful resolution consist of millions of pixels and hundreds of different diagnoses are conceivable – for each variant, many images are needed to train the neural network. To successfully train and use AI for these cases, some serious computing power is needed. Sufficient computing power and memory to use neural networks meaningfully for many real-world tasks has only been an option over the last few years.

In the meantime, AI spent a long time in what some called a winter period. Yet, that time was not wasted at all. Many methods that are still at the heart of many problem-solving strategies in AI were developed. Convolution neuronal networks (CNNs) have been established as a tool whose function is intuitively understandable in image processing. You wouldn’t inspect millions of pixels individually to recognize a holiday photo. However, if you compress the image to just a few pixels, and what remains is blue above yellow, you are probably looking at a beautiful beach shot. Long-term short-term memory models provided similar breakthroughs for processing speech. Most of this progress was barely noticed by the public, but again accelerated over the following years, giving way to many more innovations and possibilities.