Over-parameterized models tend to overfit in robust recovery tasks, therefore require carefully fine-tuning early termination conditions.
We provide a double over-parameterization method for robust natural image recovery, which does not require fine-tuning and is easy to use.
[Github]
We identify central principles in neural network architectural design that enable effective training of very deep vanilla models. Such principles may help people design new simple networks with much improved performance.
Subspace clustering is the problem of clustering a set of data drawn from multiple low-dimensional subspaces into their respective subspaces.
This python package provides implementation of scalable and provably correct subspace clustering methods that can handle 1M data points.
[Github]
Outliers are the points that do not lie in the underlying low-dimensional subspaces, which need to be detected and rejected.
This matlab code implements a random walk approach for outlier detection that can handle multiple (and possibly unknown number of) inlier groups.
[Download]
Modern datasets usually contain drastically different number of samples from different classes. which can compromise the performance of existing learning methods by a significant amount.
This matlab code implements an efficient greedy algorithm for generating a balanced subset from imbalanced data.
[Download]
