Unsupervised representation learning for image recognition – UnLIR

The proposed project lies in the field of computer vision and deep learning.
We particularly study image classification and retrieval.

As for machine learning, computer vision has witnessed a core change with the recent re-popularization of Deep Neural Networks (DNN).
Even though the recent deep learning methods offer great performance on numerous problems, several limitations are to be studied.

1) Complex tasks, such as fine grained classification problems, where better objective functions and more discriminative intermediate representation outperform DNNs.
2) The requirement for large amounts of annotated data.
3) The explainability of networks.

This project seeks answers to three questions:
-How to build networks capable of solving fine grained visual recognition tasks?
-How to train a network with few to no labeled data?
-How to explain a network decision?

More precisely, we want to improve methods capabilities to learn from few to no data in order to build highly performing discriminative representations that can address complex recognition problems. Following the methods developed, we will provide insight on how such models take their decisions.