search

cnichkawde / MatchingNetwork

Implementation of "Matching Networks for One Shot Learning" in Keras https://arxiv.org/abs/1606.04080
54 stars 21 forks source link

CIFAR10 Experiment #1

Open pGit1 opened 6 years ago

pGit1 commented 6 years ago

How would this run on CIFAR if you only trained on 100 samples per class? What Test set accuracy do you get?

pGit1 commented 6 years ago

Any comment? Can you show an example on cifar?

pGit1 commented 6 years ago

ping @cnichkawde

cnichkawde commented 6 years ago

One shot learning problem involves learning from only one example per class. The paper used Omniglot dataset which has a large number of classes with few examples per class. CIFAR10 is not a good case for one shot learning experiment as there are only 10 classes and 6000 images per class. You can certainly try with 100 examples class but that would be 100 shot learning. Also, the problem is posed as n-shot m-way which means you are presented n examples per class and the classifier has to choose between one of the m classes.

pGit1 commented 6 years ago

Can you post some code that attempts 50, 100, 200, shot learning on CIFAR? I would be very interested in experimenting with this but I dont see a clear way to do it with your code.

Thanks for your help by the way! Curious on your thoughts on this!

My thought is more concerning with few shot generalization rather than "one shot" per say. For instance can 100 shot learning on CIFAR achieve 90% accuracy on the Test set? If so then matching networks is a paradigm that lets us train any classification problem with far fewer samples.

pGit1 commented 6 years ago

ping @cnichkawde