ML Work

Machine Learning Work¶

Back in 2022 I took an introduction to Machine Learning with the wonderful Karen Mazidi who gave us a large overview of both data science basics, and basic machine learning. The class was project based, with a focus on providing documentation of the process.

Learning in R¶

We covered using R for a variety of algorithms:

• Linear Regression
• Logistic Regression
• Naive Bayes
• kNN
• k-means Clustering
• Decision Trees and Random Forests
• Support Vector Machines

As well as best practices for picking good attributes for analysis, cleaning up data in CSVs, and visualizing the results

Learning in Python¶

We then pivoted to implementing these solutions in the Python package ecosystem, using:

• NumPy
• Pandas
• Scikit-Learn
• Seaborn

To implement all the algorithms we just learned. Python allowed us to branch into Neural Networks using Keras for the implementation of zero-shot classification

Continued Work¶

This class carried into my work in NLP, which happened to coincide with the emergence of Open-AI’s ChatGPT. Check it out for work on more advanced neural nets

The Projects¶

• Low Level Basics: I worked a little bit in C++ to implement basic statistics calculations, just to make sure I had the swing of things.
• Linear Models: Assuming that problem, its input and output, are linearly related, there are multiple ways to create a predictive supervised model:
  • Linear Regression
  • Linear Classification: Naive Bayes and Logistic Regression
  • Building From Scratch
• Similarities: If instead of predicting a target value, we just wanted to understand the data, we have many different methods of breaking down complex (and high dimensional) data. This works hand in hand with dimensionality reduction. I have more on the subject here. kNN and K-means algorithms were both good tools to improve the performance of our previous models.
  • Using similarties to improve regression
  • Using similarities to improve classification
  • Clustering Spotify genres
  • Failed dimensionality reduction of Spotify
• Support Vector Machines: SVM’s divide data in such a way that optimizes the margin between data. While we can perhaps visualize data being split by a line for classification, this method can not only be used for high dimensional classification but regression as well. I have a much better description here.
  • Classifciaton with SVMs
  • Regression with SVMs
• Neural Networks: While I explore neural networks in full in my future Natural Language Processing class, we still got some experience with how neural networks work with keras and tensorflow. Using RNN, CNN, and even finetuning on Google’s MobileNet V2, I was able to create a pretty good rice identification model.

Out of anything I would recommend reading my Keras Image Classification paper for a good picture of the progress made in this class. Needless to say this was maybe the second most impactful class of my degree, right before NLP.