MNIST

• Data science pipeline applied to a classic machine learning benchmark dataset--MNIST handwritten digits

Exploratory Data Analysis

Averaged representations

Outliers extracted using matrix-matrix distances from the mean

Hyperparameter tuning the neural network

Model Results

• The confusion matrix shows accuracy prediction for each digit, and which digits are most commonly mistaken for others
• The F1 multiclass weighted score is 98.8
• 3 is most commonly mispredicted as 5 and 4 is most commonly mispredicted as 9

Metadata

• Scripts log all the important machine learning metadata in json format

Usage

• Installing the pip package python-mnist puts the data downloading script emnist_get_data.sh in your python bin directory. e.g ls PYTHON_BIN_DIR | grep mnist
• Run tests with pytest tests/*
• View json metadata with jq utility jq . models/*.json
• Project plan document found in plan.md