Pruning Neural Networks with Lottery Tickets in a MDP Approach
Suggestions for further work (possibly outside the scope of this course):
- Do you plan to compare the performance of a pruned network with training a similarly structured network from scratch?
- Use some kind of generalization for the policy (e.g. linear, NN, etc)
Stylistic comments:
- Avoid passive voice
- Keep the abstract to the point, focusing on the 4 key aspects I suggest here
- Additionally, moreover, furthermore are often (but not always) an indication of poor structure, get rid of them (e.g. Moreover, MDPs are automatic planning algorithms, which means that we are trying to find a policy that is a universal problem solver). If you have multiple points to make, enumerate them and tell the reader at the outset what you want.
Questions:
- Why did you train the LeNet first?
- Why not include actions to re-add features to the mask?