Researchers have applied the tools of neuroscience to study when and how an artificial neural network can overcome bias in a dataset. They found that data diversity, not dataset size, is key and that the emergence of certain types of neurons during training plays a major role in how well a neural network is able to overcome dataset bias.