- MNIST Handwriting
- SVM vs ANN in R with kerasa
a priori: Discuss the potential relative strengths and weaknesses of ANNs and SVMs for the dataset. Discuss the history of the dataset and the state of the art in computer vision today.
- Implement the methods (in R with Keras API for linear algebra)
- ANN – Simple NN (a couple layers, with gradient descent and backprop)
- SVM – Simple SVC
- Analyze the performance of our implementations
- Compare and contrast our two implementations with known good implementations in keras
- Write up and discuss our analysis and conclusions with:
- confusion tables, performance tables, ROC curves
- discussion of the algorithms and challenges during our implementation, potential future improvements
- Discuss how our implementations compared with production implementations (keras)
We will do it all in org, with a git repository to share our work and changes. The final project turned in will be an archive containing the org source file, the generated figures, a typeset pdf of the paper, and a typeset slideshow for our presentation.
- Note
- if perhaps this is biting off more than we can chew, we will instead
just on ANNs, following the steps as laid out above, and fo
The dataset we have chosen is the well known MNIST Handwriting recognition dataset, a mature and well-studied dataset originally compiled in 1998. There have since been many papers on ML algorithms that use this as their dataset, so we have a wealth of information on the problem at hand that we can refer to during our implementation process. This dataset is essentially the benchmark for computer vision systems in the modern world.