Kolmogorov Complexity of Kernels in CNN
This project studies the compressibility of convolutional neural networks and its relationship with Kolmogorov complexity. Final project of IA225 - Algorithmic Information and Artificial Intelligence - Télécom Paris.
This experiment aimed to investigate if convolutional neural networks, after being trained, are compressible. To do this, we applied compression and pruning techniques and verified that, without a significant loss in accuracy, a CNN can be significantly compressed and their kernel matrices have visually appealing patterns.
The report of the project can be found at the course proceedings along with the projects of my other classmates. This was the final project of IA225 - Algorithmic Information and Artificial Intelligence at Télécom Paris.
This course also has a MOOC version that is available here. I found this course to be very intresting as it gave me new perspectives on AI, mathematics and even human intelligence. I highly recommend it. The professor in charge of the course, Dr. Jean-Louis Dessalles, has valuable insights over these subjects.
This project was a great opportunity to gather experience in:
- Algorithmic information theory.
- Artificial Intelligence