This project purpose is to reproduce some of the results in Measuring the Intrinsic Dimension of Objective Landscapes and then:
- report on their main properties compared to standard smaller neural network
- test experimentally their performance for different subspace dimensions
- test experimentally their performance for different families of P
To see the resuts of the experiments look at the commented lines in runs.sh
.
The data necessary for training will be downloaded automatically with the first
run and then read from the filesystem.
A more detailed description of the project is in report.tex
.
Stuff from the authors:
Stuff from other people:
- VITALab excellent explanation!
- A good set of slides from one of the authors.
- A study using intrinsic dimensionality.
- SlideShare
- A non particularly revealing article