Does the future belong to artificial intelligence? Can artificial intelligence solve pressing problems facing humanity? It can certainly help – if we recognize the limits of the technology and apply its strengths correctly. The trick is knowing how.
Analyzing radiological images and assisting physicians in evaluating them, recognizing cars in traffic scenarios, even distinguishing dogs from cats – artificial intelligence can do all of this brilliantly. But successes such as these are rapidly raising expectations for this technology.
“Many people see intelligence as a system that thinks for itself and is very smart,” says Hanna Lukashevich, group leader at Fraunhofer IDMT. “But they’re wrong: artificial intelligence can only do what it has been taught in advance – and therefore performs particularly well in well-controlled environments. If, on the other hand, they are applied to versatile content – such as analyzing any kind of audio files with different recording quality and diverse content – this often leads to unexpected effects.”
So is it best to stay away from artificial intelligence?
“Certainly not! However, it is essential to define the appropriate model in advance – e.g. distinguishing between dogs and cats – and to train the model accordingly. If the developer did not include zebras in the definition of the model, the artificial intelligence will not be able to recognize zebras either,” explains Lukashevich. After all, artificial intelligence is one thing above all else: a tool. And just as nobody would think of drilling a hole in a pane of glass with a wall drill, artificial intelligence must also be applied correctly. The fact that it’s not always obvious what the AI model is used for makes the situation more difficult. “The most important thing is that the system is trained with data that is representative for the application,” says Lukashevich. „If, say, a model has only been trained with high-quality audio data, it will later struggle with telephone-quality audio data.“
What if the artificial intelligence just won´t do what you want?
Then it’s time to adapt the training data. “AI isn’t magic; it’s mostly mathematics,” Lukashevich assures us. “To put it another way, there‘s always a reason why something won‘t work. By and large, everything can be solved by adapting the AI components to the use case. Or have them adapted: the researchers at Fraunhofer IDMT will be happy to help.”