Curriculum Learning For Image Reconstruction Problems

The last decade has seen impressive developments in machine learning, especially in image classification. The typical work flow is to consider a set of images as training data, find the parameters of a model, e.g., a neural network or a support vector machine, and use the learned model to predict the classes of unseen/test data. When there is more than one prediction task, for example, if we want to determine if a given image depicts both cats and dogs, it is not clear whether the tasks should be carried out in parallel or sequentially. When carried out sequentially, i.e., if we determine first if the image contains a dog and then a cat, there is the potential to transfer information from the first task to the second one, and thus also an incentive to execute the easier tasks first. Curriculum learning [1] is the framework that allows determining the optimal sequence of tasks.

A different but related problem is the reconstruction of similar images (or signals in general). Two scenarios have been well studied: simultaneous reconstruction using for example, their similarity in a sparse domain (e.g., [2]), and sequential reconstruction for two signals (e.g., [3]). The problem of determining the optimal order of reconstruction when there are multiple signal has, however, never been studied.

The goal of this project is to explore how concepts from curriculum learning can be adopted in image reconstruction. It will consist of reviewing literature on both areas and conducting simulations to test new ideas. Familiarity with machine learning and optimization is desirable, but not required. Code will be implemented in Matlab, Julia, or Python.

References

[1] Y. Bengio, J. Louradour, R. Collobert, J. Weston, "Curriculum Learning", International Conference on Machine Learning (ICML), pp. 41-48, 2009.

[2] E. van den Berg, M. P. Friedlander, "Theoretical and Empirical Results for Recovery From Multiple Measurements", IEEE Transactions on Information Theory, Vol. 56, No. 5, pp. 2516-2527, 2010.

[3] J. F. C. Mota, N. Deligiannis, M. R. D. Rodrigues, "Compressed Sensing With Prior Information: Strategies, Geometry, and Bounds", IEEE Transactions on Information Theory, Vol. 63, No. 7, pp. 4472-4496, 2017.

Supervisor name: 
Dr. Joao Mota
Supervisor and Deputy email addresses: 
j.mota@hw.ac.uk

Project Type: