The goal of this project is to compare methods that calculate the intrinsic dimension of data.
Intrinsic dimension is a good characteristic of the model overfitting.
However, there are different ways to calculate intrinsic dimensionality. <Vittorio Erba, Marco Gherardi & Pietro Rotondo “Intrinsic dimension estimation for locally undersampled data” Nature>
Known methods calculate it using data samples.
In this project, we assume that the data lies in a low-dimensional data manifold.
By implicit function theorem we can describe the data dimension by rank of the functions vanishes on the data.
In the project we propose to calculate intrinsic dimension by building functions that have zero value on data and compare existing methods with the proposed one by the same procedure described in <Vittorio Erba, Marco Gherardi & Pietro Rotondo “Intrinsic dimension estimation for locally undersampled data” Nature>
We propose to investigate for which problems CIFAR is used as a training set and analyse the state of the art solution.
For this project the student is supposed to make an overview of the problems that are solved on CIFAR, such that attribute classification, face generation, person identification and others.
The goal of the project is to understand what problems are solved and how they solved, and to create library with state of the art solutions on CIFAR.