Gauging how hard a project is, is important in

  • 💸 setting freelancer rates
  • 🤝 communicating with colleagues
  • 📈 and stakeholders

Here's the system I use to classify ML tasks throughout my PhD at DTU and now at ECMWF.

🆓 Safecrackers rule

What's the first thing you do, when you crack a safe? You check if it's open.

For a machine learning project?

You check if it's been done before. Maybe someone posted it on @kaggle or @github.

Feasible in a company - not ideal for research.

🥉 Surrogate models

Sometimes you get lucky.

The data is fairly simple, but maybe the physics is expensive to calculate.

This is building a model of a well-understood process. That means we get great ground truth data and control.

🤖 The advantage is that it's faster!

🥈 Human out of the Loop

Use a machine learning model to perform a task that previously needed human intervention.

Example: Analysing x-ray images for specific diseases. Not possible for classic computers and needs radiologists.

🤖 Advantage is freeing up human 🧠 power.

🥇 Do the Impossible

Machine Learning has enabled applications that simply weren't possible.

DeepMind built a model to predict the weather in a few hours to the satisfaction of Met Office scientists!

What else is possible? Earthquake prediction? Folding proteins? ... wait.

🏆 How difficult are they?

🆓 Go for it! Just tweaking necessary.

🥉 Works with easy models, maybe even a random forest.

🥈 Harder to do. New technologies and probably some GPUs.

🥇 Extremely hard. Needs domain expertise and ML finesse.