Thesis on Convolutional Neural Network!

After a while of inactivity from blogging in the JVL blog, I finally got new material to post something relevant to the previous publication on Convolutional Neural Networks (CNN). The previous post presented a small literature review on CNN; this time I present a thesis project on the same topic.

If the material of the previous post caught your attention, then this thesis will certainly sparkle further interest. In trying to further strengthen my knowledge on CNN, I decided to take a thesis topic on Resource-aware CNN exploration on CPU-only implementations for image classification that was offered to me. This topic evidently offers that additional research and practical component that would well complement my acquired theoretical understanding. It offers a reinforcement to my learned skills. Furthermore, this project would enable me to work with different tools, libraries, and open-source projects related to machine learning and image processing. I was also looking forward to improving my coding skills in C++, especially when forking from open-source projects.

I must also say, I had quite a difficult time deciding between this CNN topic and another topic that I also found very interesting and was qualified for. The alternative was to work on a faster error correcting scheme for PUF's (Physically Unclonable Fnctions) and to implement the algorithm in an FPGA (Field Programmable Gate Array). This topic actually embodies many of my main interests: information theory, channel coding, scripting and simulations, VHDL implementation on FPGA's, and Embedded Security. As you could readily infer, it was quite a difficult choice to choose the CNN topic over this alternative.

The Thesis

The document presented here is the final document as a result of this project. This is put together with as much relevant details as possible with the purpose of making the experiment repeatable for the reader. Hopefully, the simulation benchmark results and some of the ideas exploration for reducing computation of the CNN algorithm may actually help others.

Closing lines...

After finishing this project, I feel substantially more competent in this area of machine learning. The research and exploration experience from this project gave me an opportunity to exercise as a research engineer, to benefit from and to contribute to open-source, and to compose a document that is useful to the reader, especially for sharing results and for repeating the experiment. The even more valuable take-away skills were the use of several Latex packages to produce a professional-looking documents and the improvement on coding skills and styles.

As is (generally) common to human nature, even after finishing my thesis, I still sometimes wonder "what if" I had taken the other route I described above. Where could that have taken me? What could have gone differently?