From detecting cats in images to piloting a self-driving car, neural networks can learn to handle a wide range of issues. But it’s still unclear if these sophisticated pattern-recognition algorithms comprehend the tasks they’re executing. For example, rather than learning to recognize lanes and focus on the road’s horizon, a neural network tasked with maintaining a self-driving car in its lane might learn to do so by studying the shrubs on the side of the road. MIT researchers have demonstrated that a certain sort of neural network may understand the true cause-and-effect structure of the navigation task it is being trained to execute.
Source: https://news.mit.edu/2021/cause-effect-neural-networks-1014
