This is the first project I ever wrote in Haskell, after spending a month or so getting a feel for the language. Developed as part of the assessment for an advanced topic in Functional Programming at Griffith University, as part of the requirements for a Bachelor of Information Technology (Hons). This project served as an introduction to Haskell, to the Functional Programming paradigm, and to give me significant insight into the workings of simple artificial neural networks (nothing like building something from scratch to get a good idea of how it works!).
The build will create the written component of the assessment for the project, however the code is well commented and I've tried to be as thorough as possible explaining things.
This project requires the following in order to build:
- GHC 8.02
- Latex
- Python 2.8 (for latex minted package)
adhttps://hackage.haskell.org/package/ad This has only been tested on a Linux machine running Ubuntu 16.04 however there is no reason why it cannot be built and compiled on Linux, windows and mac.
The neural network trains in a reasonable time for small mathematical
functions, unfortunately even on a tiny image classification data set (60 x 1KB
images), the performance is atrocious. This was due in part to my use of the
ad library, much faster performance could be obtained by calculating the
derivatives manually. The other part was due to my inexperience with reasoning
about the run-time performance of lazy data structures, resulting in a large
amount of GC pause time.
please contact me at donovan.crichton@griffithuni.edu.au if you have any questions, concerns, or just want to say hi.