Few-Shot Image Classification with Meta-Learning

You don’t always have enough images to train a deep neural network. Here is how you can teach your model to learn quickly from a few examples.

Why do we care about Few-Shot Learning?

In 1980, Kunihiko Fukushima developed the first convolutional neural networks. Since then, thanks to increasing computing capabilities and huge efforts from the machine learning community, deep learning algorithms have never ceased to improve their performances on tasks related to computer vision. In 2015, Kaiming He and his team at Microsoft reported that their model performed better than humans at classifying images from ImageNet. At that point, one could argue that computers became better than us at harnessing billions of images to solve a specific task. Hurrah!

However, if you are not Google or Facebook, you won’t always be able to build a dataset with that many images. When you work in computer vision, you sometimes have to classify images with only one or two examples per label. At this game, humans are still to be beaten. Show a picture of an elephant to a baby and they will never fail to recognize an elephant from now on. If you do the same thing with your Resnet50, you might get disappointed by the result. This problem of learning from few examples is called few-shot learning.

For a few years now, the few-shot learning problem has drawn a lot of attention in the research community, and a lot of elegant solutions have been developed. The most popular solutions right now use meta-learning, or in three words: learning to learn.

….

Read the full article here if you want to know what it is and how it works for few-shot image classification.