Challenges of AI

Artificial intelligence is only as good as the data it is trained with. Systematic biases in the datacan have dramatic consequences. The US Army planned to use neural networks for automatic detection of camouflaged tanks, but the network did not perform better than chance under real conditions. Due to a data bias, the network learned to distinguish between sunny and cloudy days instead of between camouflaged tanks and forests. AI systems can be deliberately manipulated, e.g., through adversarial examples or patches, which can impair the performance of image classifiers and thus pose a serious problem for autonomous vehicles. Reproducibility crisis, alchemy, and black-box problem can negatively affect the reliability and trustworthiness of AI systems. Gary Marcus points out ten weaknesses of current deep learning, including problems with transfer learning and data inefficiency. Neuroscience can help overcome these weaknesses by providing inspiration for new architectures and supporting the development of improved learning methods.