Meta-Learning

By Radek Osmulski

Learning Machine Learning

How to Fast.AI

  1. Start with the video lesson.
  2. Then head to the book chapter and run the code in the notebook.
    • Focus on understanding what’s going on in each cell.
  3. Try and recreate the pipeline from the notebook or video.
    • You can refer back for help but try to do this using your notes, documentation and whatever else.
  4. Apply what you have learned to a new dataset.

Becoming a Developer

Make use of long, uninterrupted work sessions.

Flow

Good developers spend a lot of time studying their editors and swear by the ones that don’t require you to lift a finger. When you reach for you mouse your flow is interrupted.

How to Learn

Make your learning active.

The Hidden Game of ML

The hidden game is the ability to generalize to unseen data.

How to Structure a Machine Learning Project

  1. Create a good train-validation split.
  2. Create a baseline.
    • The simplest transformation of the training data that will move the needle on the metric measured by the validation set.
    • Helps to better understand the problem.
    • Helps remove bugs from the pipeline.
  3. Next, work on the part of the pipeline that will benefit most from tweaking.
  4. Invest time early in exploring a larger set of architectures and developing diagnostic code.
  5. Move in small increments.
  6. Consider constructing a smaller train set so you can test if you changes are having a positive impact more quickly.

How to Win at Kaggle

  1. Join a competition early.
    • Create a baseline.
  2. Read the forums daily.
    • This will take you 80% of the way there.
    • It’s the only way to keep up with developments.
    • Make posts about papers you think are relevant to the competition.
  3. Make small improvements every day.
    • Focus on small teaks that will compound over the course of the competition.
    • Combine predictions form multiple models and in the hopes their errors are not correlated.
    • Cross-validation is used in most good Kaggle submissions.
  4. Find a validation split that tracks the leader boards.
    • If you are doing similar things to others, you should be getting similar results.
    • If you changes in the leader board are not reflecting local changes, revisit you validation set.
  5. Posts from top Kagglers will get you 80% of the way.
  6. Use research papers, blog posts and creativity for the last 20%.
  7. Ensemble results.

The Best Hardware for Deep Learning

What GPU should you get? The biggest (most RAM) you can afford.

On Finding a Job

Try and meet people who can influence hiring decisions where they hangout.

Credibility

This is the name of the game in machine learning. You build credibility by being visible online, share your work and be helpful on forums.

The Party is On Twitter

Twitter is where the deep learning community hangs out. Twitter is a great way to stay up to day with research and trends.

Sharing Your Work

You should start sharing your work as soon as possible. You are in a unique position at any stage in your learning to be able to communicate concepts well to other people at your stage of understanding. Writing and reflecting also helps you to understand what you are doing on a deeper level.

What to Focus On

Speak to your experience. This is one fot he most valuable things you have to offer as others can see what you did an how as well as how you felt along the way and what results you achieved.

Find a Mentor

A good mentor is someone very good at something you care about. A mentor doesn’t even need to know you or even be aware that you exist.

Permissionless Apprenticeship

You receive value by giving value first.

Another way to attract the attention of like minded people is to Learn in Public.

The Regrets of fast.ai Students

Not spending enough time doing application. Aim to spend 80% of your time doing and 20% of your time learning content.

Tap in to the Power of Community

Start asking question on the fast.ai forums. Overtime, you will start to see the value of these exchanges. Insights from others.

Energize

When and what you eat matters A good way to complete more mental work is to eat less.

Side Notes