Text

Make any machine learning algorithm deep

Jeroen Tempels

Text

Limitations of shallow learning

k-layers -> compact representation

(k-1)-layers -> exponentially larger representation

Text

Text

The concept of deep learning

Using k-means for deep learning

1) Select random patches from image

2) Cluster the patches using k-means

3) Use centroids to construct features

K-means

K-means

1) Select random patches from image

3) Use centroids to construct features

General thesis idea

2) Cluster the patches using any clustering algorithm

Any clustering algorithm

Any clustering algorithm

Automated deep hierarchy construction

1) When to stop adding layers?

2) Amount of clusters at each layer?

Text

Approach

1) Implement k-means pipeline

2) Test on CIFAR 10 dataset until comparable results

3) Try different clustering algorithms

4) Find criteria to automatically construct deep hierarchy