Note on the Reliability of Biological vs. Artificial Neural Networks

Various types of neural networks are currently widely used in diverse technical applications, not least because neural networks are known to be able to "generalize." The latter property raises expectations that they should be able to handle unexpected situations with similar success than humans. Using fundamental examples, we show that in situations for which they have not been trained, artificial approaches tend to run into substantial problems, which highlights a deficit in comparisons to human abilities. For this problem-which seems to have obtained little attention so far-we provide a first analysis, based on simple examples, which exhibits some key features responsible for the difference between human and artificial intelligence.