Active perception and disentangled representations allow continual, episodic zero and few-shot learning | ScienceToStartup | ScienceToStartup