I would like to learn more about machine learning and build something interesting in my free time. Problem is I am so done with the online courses, because those classes requires you to take the bundle of courses and to learn something I might have known before. As a result, I looked up the discussions from reddit and found a curriculum here The goal of the curriculum is to learn how to build and deploy a ML application in the end. Since most of them are reading materials, I could pick up what I want and tailer those resource into my own learning path.

My goal:

• I would like to learn more about machine learning, deep learning, and then GenAI.
• Other than building projects in the journey, I would like to deploy them on the web.
• Also, I am interested in visualize the data and explain what the model tells us.

My background:

• I’ve learned general AI and machine learning concepts in grad school, but I did not have more in-depth knowledges in deep learning.
• When I was taking machine learning in signal processing, I’ve learned PCA, ICA to reduce dimension. So, I’ll skip them.

Routine:

The schedule each week would be as followed. Day 1 - 4: Reading material, lesson and exercise implementation Day 5 - 6: Assignment implementation Day 7: Document, share and self-reflection on my plan/goal.

Adjusted resource:

Milestone 1: Machine Learning

Since I’ve gained knowledges in python and mathematics, I will start from week 3 from LearnML and extend my learning from kaggle learn

week 1 - 4: Data visualization

Some topics that interst me are:

• Intermediate Machine Learning: est. 4 hours
• Data Visualization: est. 1 hours
  • I just want to review different method to explain the data. So, I only need to read the tutorial in summarized tutorial
• Feature Engineering: est. 5 hours
• Machine Learning Explainability: est. 4 hours

Reflections:

• The kaggle material are easy to follow and they did not put too much challenges for me. I will need to adjust the plans in month 2 to get more challengs and implement more projects.
• I want to make the plan as milestones instead of month, because I felt more comfortable when there are some other commitments to focus on.

Milestone 2: Deep Learning

The overlapped materials people talks about are:

• fastAI’s course Practical Deep Learning
• a text book Dive into Deep Learning. FastAI’s course looks more interesting to me, so I will explore FastAI’s course and plan to learn more from this material and make the materials from the curriculum as supplement resources.

Milestone 2.1.1: Neural Networks

• Practical Deep Learning(Part 1): est. 14 hours
• Dive into Deep Learning(Chapter 3. - Chapter 9.): est. 10 hours
• Assignment: on github.

Reflection for Milestone 2.1:

It is unrealistic to complete part 1 in 14 hours. because the steps to get full reward from the class is to watch the video, reproduce the notebookm, and apply the model on different dataset or do the additional assignment. I only completed lesson 1-3 and realtead reading. I will adjust the plan to work on lesson, reproduce the result or implementation.

Milestone 2.1.2: Neural Networks

• Videos: Lesson 5
• Textbook:
  • 04_mnist_basics
  • 05_pet_breeds
  • 06_multicat
  • 07_sizing_and_tta

• Project: Feedforward NN for retail dataset (Colab notebook).

• Small Tasks:
  • Run 04_mnist_basics.ipynb in Colab
  • Watch lesson 5
  • Run 05_pet_breeds notebook and modify dataset
  • Try 06_multicat for multi-label example
  • Apply augmentation tricks from 07_sizing_and_tta
  • Build feedforward NN for retail dataset and upload to Colab

Milestone 2.2: Transformers

Milestone 2.2.1: FastAI

• Videos: Lesson 4 (NLP)
• Textbook:
  • 10_nlp
  • 12_nlp_dive
• Small tasks
  • Watch Lesson 4 (NLP) video
  • Run 10_nlp.ipynb
  • Run 12_nlp_dive.ipynb
  • Modify the notebook: Use your own small dataset (e.g., sentiment analysis of tweets or movie reviews)
  • Train a small model
    • Make predictions on a test set.
    • Check metrics (accuracy, F1, etc.).
  • Optional mini-project: Quick Colab notebook of text classifier ready to share or save.

Milestone 2.2.2: Hugging Face course (Transformers & Fine-tuning)

• curriculum
• Small Tasks:
  • Follow Hugging Face Course basics
    • Load pre-trained models and tokenizers.
    • Fine-tune on a small text dataset.
  • Implement a simple chatbot: Use a Mini-GPT model or small GPT2 variant.
  • Test the model
    • Have short conversations, log responses.
  • Optional Deployment

Milestone 2.3: Diffusion(FastAI)

• Videos: Lesson 8 (CNNs), Lesson 9 (Stable Diffusion)
• Textbook Chapters:
  • 13_convolutions → CNN fundamentals
  • 14_resnet → Residual networks
  • 15_arch_details → Architecture details
  • 16_accel_sgd → Accelerated SGD training methods (optional, helps with large models)
• Small Tasks:
  • Watch Lesson 8 (CNNs)
  • Run 13_convolutions & 14_resnet in Colab
  • Watch Lesson 9 (Stable Diffusion)
  • Implement a small generative image project using a pre-trained diffusion model
  • Deploy a design generator via Hugging Face Spaces

Milestone 2.4: The Rest of fastai

• Videos: Lesson 6 (Random Forests), Lesson 7 (Collaborative Filtering)
• Textbook Chapters:
  • 08_collab → Collaborative filtering
  • 09_tabular → Tabular models
  • 11_midlevel_data → Mid-level data API
• Small Tasks:
  • Watch deployment & ML breadth videos
  • Run 08_collab notebook → build recommender
  • Run 09_tabular notebook → tabular prediction
  • Explore mid-level API via 11_midlevel_data
  • Build small applied ML project (e.g., recommender or tabular Kaggle dataset)

Milestone 2.5: Deep Reinforcement Learning(Simonini Thomas)

• Follow syllabus Deep Reinforcement Learning Course(8 Units): est. 16 hours lesson and 20+ hours implementation
• Assignment: on github.

Milestone 3: GenAI/MLOps (est. TBD)

• Kaggle 5-Day Gen AI Intensive Course

reference resources:

• LearnML

Bao-Jhih Shao

A software engineer writing something to keep the memory.

Comments

  Write a comment ...