#9 HandWritten Digit Recognition

---

I was in sort of a rut in the past few months, I was trying to work on projects that were clearly above my existing skillset. It took some time before I realized to cut my losses and work again on simpler and shorter projects. While admittedly, I didn’t learn a whole lot new during this project, It did serve as a good refresher.

---

Data Sources Used

Digits database that comes inbuilt into Scikit It contains (1797) images of labeled handwritten digits made by Gael Varoquaux Loaded using the following code.

Methodology

Part One: Libraries Used

Part Two: Setting Up Dataframe

I used the load_digits() function imported from the load_digits library to set up the dataframe. In order to preview the images and their corresponding labels, I used the algorithm provided by Michael Galarnyk. From what I understand, it works like this.

1. create a blank fig.
2. run a loop calling the first 16 entries in digits.data, digits.target as "image" and "label" respectively.
3. create a subplot inside the blank figure of 2 rows and 8 cols.
4. place images in them (8,8) size and monochromatic color.

Part Three: Train Test Split

Part Four: Testing Algorithms

Here, I have streamlined the process of testing an ML algorithm on a clean dataset. It’s a simple four-step process.

Part Five: Accuracy Results

Logistic Regression: 0.9666666666666667
Support Vector Classifier: 0.9844444444444445
Random Forests: 0.8066666666666666
Knearest neighbors: 0.9733333333333334

Apart from Random Forest Classifier, every other model seems to work out. Maybe I’ll update the post if I find a more efficient configuration for it.

Part Six: Heatmaps

Going through the heatmap of Random forest, it’s clear to see it’s struggling with recognizing certain digits. especially, 1, and 8. – Where the model faltered the most.

Like always, The full code in its entirety can be found here.

---

REVIEW

As stated in the intro, this project was more of a refresher to get me out of a rut. As such, I refrained from implementing algorithms and libraries I was unfamiliar with. I did manage to revise the following.

1. Using the four-step process of machine learning.
2. Algorithm to plot images from dataframes.
3. Utilizing Scikit’s Make_pipeline library for model instancing.
4. Using seaborn for heatmaps.

---

RESOURCES

https://towardsdatascience.com/logistic-regression-using-python-sklearn-numpy-mnist-handwriting-recognition-matplotlib-a6b31e2b166a

https://scikit-learn.org/stable/auto_examples/classification/plot_digits_classification.html