A simple custom digit recognizer implementation.
- Create 4 models for multi-class classification of the input images
- A, B) Two Basic Machine Learning Models
- C) Multi-layer Perceptron
- D) Simple Convolutional Network
- Stand-alone Jupyter Notebook
- Solutions Implemented with Sklearn, Tensorflow, and/or Pytorch (free choice by each student or mini-group)
- Understanding and EDA (Exploratory Data Analysis) of the dataset provided
- Build the Models
- Train the Models
- Cross-Validation + Basic GridSearch hyper-parameter tuning
- Performance report / Comparison of results
- Code (documented and executed)
- Models
- Selected model (ready to be scored independently with a “test” set)
As defined in the attached rubrics file, the grading will have 3 components:
- 35% CONCEPTUAL
- 35% CODING
- 30% BAKE-OFF Ranking of results (performance)
For the scoring and ranking of results, an IID(*) “test” set will be used.
Note: the ranking grade is based on the “discrepancy” of performance with respect to the average standard solution (only really under-performing models would have a significant penalty in this rubric).