What might be missing for high performance one-shot learning models? CNNs need atleast hundreds if not thousands of samples before one could classify or detect things with near-human accuracy using say fine-tuning of a pre-trained network on Imagenet. What are we missing before we can have a one-shot learning system, where say something new like a fidget spinner is invented in the world and we can get it identified or detected after showing just 1 image (like children do). Does something radical other than convolutional nets needs to be created or are we just waiting for 'that magical architecture' to be discovered by trial and error on arxiv or is it something else.
 A: Well, there are many techniques aimed at solving few-shot learning problems, especially in the realm of computer vision. CNNs, by themselves, are not designed to solve such problems. However, their design  is suitable to be used in techniques aimed at solving few-shot learning problems. For example, they can be used as part of an unsupervised model that learns good representations of the data, i.e, pre-training as you stated. Hopefully the model learns how to represent objects somewhat generally, and it doesn't just learn how to represent objects of very similar classes to those seen in the data. 
Now, in some ways, pre-training can be considered meta-learning. And in the realm of meta-learning, a very promising avenue for the few-shot learning problem is model agnostic meta-learning. To me, the performance of MAML techniques (among many others) show that there may not be that much missing from solving one-shot and few-shot learning problems to begin with, at least when it comes to computer vision.
https://arxiv.org/pdf/1703.03400.pdf
