About

Why I keep my deep learning notes as a website, how the notebooks are organised, and how to run them yourself.

About

Notes I can read on my phone

A Jupyter notebook is a good place to learn and a poor place to revise. So I publish mine as a website: the same code and output, readable anywhere, in the order I’d want to relearn it.

Who writes this

I’m Arun Koundinya Parasa, and I work with data for a living. I started these notes in 2024, during my M.S. in Information Science (Machine Learning) at the University of Arizona. The Amazon reviews project was a course project there, done with Patricia HoiKei Woo.

How the notes are organised

The notes follow a learning path of four tracks:

  1. Tensor foundations: tensors, shapes, broadcasting, the functions I kept reaching for, and gradients with GradientTape.
  2. First neural networks: linear regression as a one-layer network, and an honest look at when more layers don’t help.
  3. Projects: sentiment classification of Amazon reviews with an LSTM, and vegetable image classification with CNNs and transfer learning.
  4. Language and embeddings: Word2Vec, the Hugging Face pipeline and fine-tuning, and BERT and sentence-transformer embeddings.

The home page shows the whole path, and each note links to the one before and after it.

Run them yourself

Most notes are Jupyter notebooks. Use the Open in Colab button at the top of a note to run it in your browser with a free GPU. The pages show the output saved when I ran them, so library versions may have moved on since. Everything is on GitHub. Copy whatever helps; a link back is appreciated.

Comments are open at the bottom of each note if you spot a mistake or know a better way.

The theory behind the code

These notes are the practical side. For the ideas behind them, explained from scratch, see my free video series AI Basics with AK. Its post on representation models is the companion to the last note here.

Elsewhere