Applications of few shot learning

1

Could someone inform me applications of few shot image classification, or few shot voice detection, I am not able to think of any, it makes it difficult to understand why do we need this.

Is one application that, a software would be able to detect persons voice by hearing it once, so Siri would be able to tell whether it is person 1’s voice or person 2’s voice?

Is one application that, we could have an app that looks at an object and is able to tell its category (but don’t we already have such apps)?

Could someone please inform me some easy to understand applications of few shot learning?

2

In my opinion one huge advantage is the data size reduction needed for few-shot learning.
Collecting and cleaning a dataset is extremely costly and often users underestimate the necessary work load to create a dataset such as ImageNet.

Another argument for few-shot learning might be that “a baby doesn’t need to see 100k images of a giraffe to recognize one”. So the argument is that the sample efficiency of a lot of DL models is bad and has to be improved.

1 Like
3

but I am still confused for what application will it have, for example,

  1. will it lead to creating some more intelligent game avatar, that is able to get used to surrounding game environment fast, as it would be able to comprehend objects around it in one shot?
  2. another title that I have seen is ‘learning to learn’, does it mean that game avatar would be able to learn to play game fast, so one could see an ai like we have for the games of dota, go, for other games also?
  3. is the facial recognition system we use on phone, already few shot, as I show it image of my face once (with some movement) at the beginning (when I use my phone for the first time), and it learns how my face looks, and now only I would be able to pass facial recognition on my phone?
  4. is this one correct? a virtual avatar would be able to detect persons voice by hearing it once, so Siri would be able to tell whether it is person 1’s voice or person 2’s voice?
4

All points could be answered with “probably yes”.
If you see what Deep Learning methods are able to do today (and especially when they fail), you could imagine to create these applications in a less amount of time and will less resources.

So e.g. for your last example: if you train a model, which would achieve a high accuracy of person identification with a single sound sample of 1 second, then yes, Siri would then be able to classify you instantly.

5

Ok, thanks for your reply.