Final Project ECBM E6040

Final Project
ECBM E6040
INSTRUCTIONS: For the final project, we have compiled a list of papers that will provide
you with an extensive exposure to applications of Deep Learning to Object Recognition.
Students can form a team of 1-4 members. Each team can pick one of the papers and will
be asked to review and recreate the results from the paper using Theano. However, more
work in terms of recreating all the results from the paper and achieving the best performance
will be expected from teams comprising total of 4 members. Submission should include the
source code, Jupyter notebook, and a report. The source code should be well-documented,
and the Jupyter notebook should demonstrate that your code is executable and your results
are reproducible.
Each group should create a private Bitbucket repository for the final project, and grant
the read access of the repository to the account E6040TA. The repository should contain all
the source code and the Jupyter notebook. Every member in the group should contribute
to the repository. Please organize the repository nicely, and provide README files.
Submit the report to Canvas. Please DO NOT put your report in the repository, and please
DO NOT submit your code to Canvas.
The final report should
• Summarize the paper and provide a brief review of the relevant literature.
• Recreate results in terms of test accuracies and all plots and figures in the paper. Also
include training time for all models you present in the report.
• Provide a comprehensive discussion of your results and compare them with the results
of the paper. Discuss whether you were able to get the same or better results than
those described in the paper. If not, provide a discussion on what could have been
done to improve your results.
• Discuss what insights you were able to gain from your implementation/experimentations.
• Summarize in detail the contributions of each team member in terms of coding and
writing the report.
We are aware of the code available online for the provided publications. Please note that
you are not allowed to reuse code available online and are expected to implement the final
project in its entirety by yourself. You may however reuse any code from the homework
assignments.
The project is due at 11:59 PM on Thursday, May 5th, 2016.
Finally, if your team would like to work on a project outside of the papers provided, you will
have to make a proposal and get the project approved from the instructors beforehand.
Please refer to the post on Piazza for information regarding group formation.
List of Papers
1. ReNet: A Recurrent Neural Network Based Alternative to Convolutional Networks
2. Deep Networks with Stochastic Depth
3. Spatial Transformer Networks
4. Striving for Simplicity: The All Convolutional Net
5. BinaryConnect: Training Deep Neural Networks with Binary Weights during Propagations
6. Recurrent Convolutional Neural Network for Object Recognition
7. Spectral Representations for Convolutional Neural Networks
8. Regularization of Neural Networks using DropConnect
9. Network in Network
10. ALL YOU NEED IS A GOOD INIT
11. Multi-Digit Number Recognition from Street ViewImagery Using Deep Convolutional
Neural Networks
12. Convolutional Neural Networks with Low-Rank Regularization
GOOD LUCK!
2