Applied Machine Learning Course
July 27, 2024
Who should take this course?
Applied Machine Learning Course is built around developing a solid foundation to understand
statistical, machine learning, and deep learning methods without delving too deeply into the
mathematics and theory behind them. The focus is on the application side of the methods to
ensure that one can independently develop industry applications.
When a developer is given a machine learning project, the typical path is this: mathemati-
cally formulate the problem, understand data availability, correctly choose the right methods,
perform out-of-sample testing, and understand the avenues for performance loss. This is an
iterative process and experimental science. However, repeated iterations and experiments can
lead to productivity loss; one gets better with experience, as they develop an intuitive and
nuanced understanding of the methods and the process involved. Very often, one also has
to consider the right choice of methods under constraints such as application response time,
production cost, and data security.
A machine learning system is built on many pillars: cloud engineering, data and analytics
platforms, and MLOps. However, our focus will be on understanding the core models that
power a machine learning system. The Applied Machine Learning Course is built around de-
veloping an intuitive understanding of statistical, machine learning, and deep learning models
with a focus on industry applications. This implies that we will also discuss how to incorpo-
rate business understanding into the model development process. We might also discuss some
analytics to measure model performance.
The course is designed for developers or engineering managers aiming to develop machine
learning systems. It should be relevant for people from all industries, and that shouldn’t be
an impediment to enrolling in the course. A project manager developing an AI feature might
also find this course useful, though to a lesser extent.
A typical course on machine learning methods involves a lot of mathematics. We are going
to deviate from that pattern. I believe that understanding intuition is more important than the
actual mathematics behind these methods because, in industry, one evaluates these methods
based on their ability to add value to a business application. With the proliferation of readily
available libraries and code, it becomes less important to have an in-depth understanding of
the mathematics behind machine learning methods. In an industry setting, it becomes more
crucial to understand the reasons for lower accuracy for a given method. This can only be
achieved if one has a good intuition about a given method and has a sharp grasp of data
wrangling and analysis.
However, a complete aversion to mathematics is also not warranted. Having mathematical
clarity about these methods does add value, though it comes at the cost of learning time.
Hence, we will optimize for time and discuss the mathematical aspects of these methods only
when it is of utmost importance and cannot be ignored. Even then, I will try to present the
complex mathematical notions in an easily consumable manner.
1
Course Structure
This course is planned to be for 75 hours, with each class being 90 minutes on every Sunday for
50 weeks. I have purposefully chosen one session per week. This to ensure easy consumption
of the material and to have enough time for the participants to revise and work on takehome
assignments, while they grapple with their full-time job. Each class will have a maximum of
15 participants to avoid crowding. For now, I have written a course structure for the first 7
classes, and plan to update this over the next couple of days. Feel free to give suggestions, if
you want something included.
• Week 1: Fundamentals of Machine Learning
– Introduction
– Why does mathematics play a major part in Machine Learning? Major differences
between Data Science, Machine Learning, and Deep Learning.
– Understanding Data Generation Process and Parametric/Non-Parametric models.
– What led to the rise of Deep Learning Models?
– What led to the rise of Python Programming Language? And why is it the go-to
language for Machine Learning models?
– How is GenAI different from Machine Learning?
– Example AI scenarios and typical challenges faced.
• Week 2: An introduction to Large Language Models (LLMs)
– Application examples to Large Language Models.
– Conceptual Understanding of why Large Language Models became popular.
– An introduction to training sample and zero-shot learning.
– Data collection process and costs associated with it.
– An introduction to n-shot learning.
– Difference between open, open-weight, and closed LLMs.
– An introduction to the building blocks of a Large Language Model: tokenization
and token embeddings.
• Week 3: Designing Tokenizers
– Word Tokenizer
– Subword-Based Tokenizers
– Sentence Tokenizer
• Week 4: Designing Token Embeddings
– An introduction to TF-IDF
– Properties of a good token embedding
• Week 5: Feed-Forward Neural Networks
– Understanding the building blocks: input layer, hidden layers, output layer, neu-
rons, activation functions
– PyTorch vs TensorFlow
2
– Training Data, Test Data, and Validation Data
– takehome: class-prediction using feed-forward neural networks
• Week 6: An Application of Feed-Forward Neural Networks
– Discussion on the takehome. Discussion on the architecture design choices.
– Discussion on the various activation functions.
• Week 7: Back to Embeddings
– Conceptual Understanding of Word2Vec.
– Architecture for Word2Vec.
– Discussing PyTorch Code for Word2Vec.