Fair representation for safe artificial intelligence via adversarial learning of unbiased information bottleneck