-
Cover story
Learning machines
It turns off the light, translates texts and suggests bargains when shopping online: artificial intelligence has long since taken its place in everyday life. It is also developing into an important basis in science. For although machines are still a long away from real intelligence, they are clearly superior to humans in some areas – for example in the recognition of hidden patterns.
The waters hit the place without warning. On 29 May 2016, a thunderstorm cell erupted over Braunsbach in a very short time. As much rain fell within one hour as would normally in months. A huge flood wave made its way through the village and swept away everything that stood in its way: trees, cars, walls of houses. It left behind about 50,000 tons of rubble in the village northeast of Schwäbisch Hall. Total damage: over € 100 million.
Even though the weather forecast has become more accurate over the years – it is still difficult for meteorologists to issue timely warnings of heavy rain or local thunderstorm cells for specific places such as Braunsbach. This is partly due to the relatively coarse resolution of the German Weather Service’s (DWD) regional weather models. “Anything smaller than two kilometres will fall through the grid. The model will then say, for example, that it rains in an area measuring two by two kilometres – even if in reality blue skies and rain alternate in the area. It is usually not enough to reliably predict precipitation locally,” explains Dr. Martin Schultz.
In the DeepRain project, the physicist from the Jülich Supercomputing Centre is therefore trying to improve the forecasts so that the authorities will have sufficient time to warn of local thunderstorms and heavy rain. Artificial intelligence (AI) is to make this possible. It is to search for patterns in weather data that announce local weather extremes.
Rule-based calculation
AI is an approach to replicate intelligent behaviour using computers. The machines learn, draw conclusions and correct themselves. However, they are still far from reaching the human brain. Our brain is more energy-efficient than any other machine; it can draw meaningful conclusions from just a few examples; it is particularly capable of thinking flexibly, finding unconventional solutions and drawing connections between completely diverse facts.
Machines, on the other hand, have an advantage when it comes to stoically ploughing through mountains of data and tracking down hidden patterns or recognising much more complex patterns in a mass of information than humans can.
Classical face recognition methods are a common example: AI sorts the photos of different people by, for example, pupillary distance, face shape and nose size – depending on what the programmers have specified. It then creates a pattern for each face. After that, it applies its knowledge to new pictures: it compares a picture with the present photo stock and suggests who can be seen on the picture. So AI is being taught to judge new data sets. This is one of the simplest forms of machine learning.
The pattern of rain
In the improved weather forecast that Martin Schultz is aiming for, the machines need to be able to do a little more to recognise patterns. “Weather data contain complex temporal and spatial patterns. We don’t know which of them are typical for heavy rain. We therefore feed the software with as much data as possible. It searches for patterns itself and then makes forecasts.”
Martin Schultz uses an advanced form of machine learning: deep learning. In this, AI systems also search large amounts of data – in the case of DeepRain, weather data from previous years. However, the researchers do not specify what is characteristic of extreme weather. Instead, they train the machine to find out for itself.
“We don’t know what patterns AI is looking for. These can be things that we haven’t even begun to think about or that we might never have recognised,” says Schultz. At the end of the day, however, he and his colleagues can check whether the AI forecast is correct, such as if it rained heavily on that day, and can then report this back to the software. Through constant repetition, AI “learns” which patterns best predict heavy rain.
The functioning of deep learning is roughly similar to the learning processes of our brain. Countless billions of nerve cells are interconnected there. In this way, they pass on information and process it. When we learn, we use certain connections between nerve cells again and again, thus changing the network between the cells: in children who read a lot, for example, the connections between the areas of the brain responsible for vision, hearing and speech are strengthened. In professional badminton players, the network of the brain regions that coordinate vision and movement changes.
Deep learning uses simple mathematical units, the activity of which roughly corresponds to that of nerve cells in the brain: they are also linked via input and output connections and receive information from other units that process and forward them. However, they function much more simply than the biological models. The mathematical units are organised in layers.
Martin Schultz would like to use AI to improve local weather forecasts.
Thousands of layers
“Deep networks for deep learning sometimes have hundreds to thousands of layers in which the data is processed,” explains Dr. Jenia Jitsev, who is working on the architecture of such models at the Jülich Supercomputing Centre. In face recognition, for example, it is as if the input image passes through a variety of filters that respond to increasingly complex patterns. The first layer only perceives brightness values. Deeper layers react to edges, contours and shapes, while even deeper layers eventually react to individual characteristics of human faces.
The network learns to identify a given face by remembering the combination of brightness values, edges, shapes and details that characterise this face: as with the nerve cells in the brain, certain connections between the network units are strengthened and weakened. The learning process creates connection patterns that lead to the correct result. “Deep neural networks need as many different training examples as possible: the more different examples it gets, the more successful the learning,” says Jitsev.
This is exactly one of the problems that Schultz still has to solve in the DeepRain project: there is a lack of training material. “We transfer 600 terabytes of data from Deutscher Wetterdienst [German meteorological service] for our calculations. That doesn’t sound like a shortage in the first place.” However: heavy rainfall is rare. “According to statistics from Deutscher Wetterdienst, there were no more than eight such events at any one station between 1996 and 2005,” says Schultz. Accordingly, data sets from which a pattern for the AI could crystallise are rare.
Furthermore, the data is not only required for training, but also for the final quality test. According to deep learning expert Jenia Jitsev, “Typically, only 80 per cent of the data is used for the training phase. The remaining 20 per cent won’t be touched at first. This test data set is retrieved only after the training in order to test the capabilities of the neural network.”
Hopes to find patterns in the brains of people with psychological and neurological diseases, using AI: Simon Eickhoff.
Jenia Jitsev deals with the architecture of deep neural networks for deep learning.
Individual forecasts
This test phase is particularly important when it comes to sensitive data – data that determines the fate of people, such as the selection process for applications, the assessment of creditworthiness or medical diagnoses. Prof. Simon Eickhoff from the Institute of Neuroscience and Medicine (INM-7) is working on the latter. Using AI, he someday hopes to find patterns in the brains of people with psychological and neurological diseases so that they can be treated specifically and individually. For example, computer programs are to search brain scans for patterns that provide information on how likely a relapse is in a patient with depression. AI could predict how quickly the impairments in a person with Parkinson’s disease will progress or whether a patient can be treated better with drug A or with drug B.
The overall pattern counts
However, there is still a long way to go. Eickhoff and his team are already working on using pattern recognition to enable AI to obtain certain information from brain scans: at the moment, the focus is on cognitive performance and personality traits such as openness, sociability and emotional stability. For this purpose, Eickhoff and his team have trained machine learning programs with the brain scans of hundreds of people. Certain psychological parameters of these test persons are also entered, such as the reaction time in a standardised test. If the model has seen enough data, it can deduce the reaction time of a new individual from the brain images alone. “However, our algorithms do not search for individual aspects in the image data. We can’t say that, in people with a good working memory, certain areas of the brain are larger than average. Rather, the overall pattern is decisive,” says Eickhoff.
“Deep neural networks need as many different examples for training as possible: the more examples, the more successful the learning.”
Jenia Jitsev
According to the brain researcher, more complex cognitive abilities, such as reaction times or the capacity of the working memory, can be deduced relatively reliably from brain scans using AI. Even though the prediction also tends to be correct in the case of personality traits, it is not that accurate yet. Quality assurance with data that AI does not yet know reveals this: it only trains with one part of a data set at a time. The researchers use the rest to check how well the AI predicts personality traits after a learning phase.
AI already provides very good results in predicting age and gender. “Here, our program can indicate with a certainty of 90 per cent whether the brain belongs to a woman or a man. Regarding age, it is in the range of plus/minus four years,” reports Eickhoff.
Bringing light into the dark
The verifiability of data such as age or gender is comparatively simple. It becomes more difficult with diagnoses and prognoses. “The acceptance of artificial intelligence in healthcare hinges on the trust placed in it – by both patients and doctors,” believes the Jülich expert. Trust is partly based on the fact that it is plausible how a diagnosis or a result comes about. However, in deep learning, AI experts like to compare a neural network to a black box: you know the input data and get an output. However, the processes in the information-processing layers in between are so complex that it is usually impossible to understand how the network arrives at its results. It is therefore an important task for the AI experts to shed some light on this darkness in the coming years and to make the complex patterns on which the results of the AI are based visible to us, says Simon Eickhoff. In this respect, many experts are hoping for “explainable AI”. In addition to the actual result, such AI also provides the criteria as to how it came to its conclusion. Not only medicine and neurosciences would benefit from such algorithms, but also weather forecasting, speech recognition and the control of autonomous cars. “Only when we can explain why an algorithm has made its decision will we accept machine-generated solutions that our brains cannot find,” says Eickhoff.
Arndt Reuning
AI on the rise
As early as 1956, the term of artificial intelligence was coined at a several-week workshop at Dartmouth College in New Hampshire. First concepts for artificial neural networks already existed at that time. In the 1970s, however, the long “AI winter” began: research stagnated due to a lack of both computing power and sufficient training data. Around the turn of the millennium, the renaissance of machine learning began.
“Only big data and progress in learning algorithms have made the current progress of artificial neural networks possible,” says deep learning expert Dr. Jenia Jitsev from the Jülich Supercomputing Centre: the more data, the more examples with which an artificial neural network can be trained. And the more intensive the training is, the better a network can correctly classify new examples it has never seen before. Such self-learning AI algorithms are, however, very computationally intensive. “Here, we benefit from the increased performance and storage capacity of modern high-performance computers, which have been specially developed to handle such algorithms and huge amounts of data.” In addition to conventional processors (CPUs), many graphics processors (GPUs) can be found in such high-performance computers. GPUs may indeed have slower cores than CPUs, but they still have a decisive advantage: unlike CPUs, they have thousands of cores that can perform simple computing operations in parallel with great efficiency. This is ideal for the functioning of neural networks and deep learning: during their training, a large number of such operations have to be carried out repeatedly.
High-quality software adds to the ever more powerful computers: “Meanwhile, many different open source tools are available.” The three prerequisites – big data, powerful hardware and suitable software have – finally awakened AI from its long hibernation.
Photos: Forschungszentrum Jülich/Sascha Kreklau, Forschungszentrum Jülich/Ralf-Uwe Limbach, Bernd Struckmeyer