Chapter 13 The T-Shaped Modeler
- The T-shaped modeler excels in some modeling mindsets and has some beginners skills in the others.
- Knowing multiple modeling mindsets means less risk of getting stuck, more creativity and more pragmatism.
13.1 Be Like The Octopus
The shark is coming for the octopus.
Flee. The octopus darts away with a jet of water. But the shark has almost caught up with the octopus.
Ink. In a dark explosion of liquid, the octopus squirts ink into the shark’s eyes. A brief window of opportunity for the octopus.
Hide. The octopus begins to squeeze into a tiny notch in the rocky reef. Too slow; the shark sinks its sharp teeth into the octopus’ squishy tentacle.
Bite. The octopus counters with its own sharp teeth.
Venom. The octopus’ bite was off and the venom is nothing more than an itch to the shark.
Drop tentacle.The octopus sacrifices the captured tentacle for a chance at survival. Escape now, regrow later. The shark gladly accepts the meal and resumes its pursuit. But where is the octopus?
Camouflage. The octopus has become a part of the reef and lives another day.
The octopus is a cunning creature. But also a squishy delicacy to humans, sharks and other predators. To survive, the octopus has many tricks. Famous for their ink and camouflage, they even have a venomous bite and cover themselves with things like rocks. Imagine an unsuspecting shark just craving octopus sashimi. Only to come face to face with an octopus paladin in freaking stone armor. The octopus has only a short life of about 1-2 years. I wonder what the earth would look like if octopuses lived as long as humans. Would they build cities and rule the world? But I am digressing.
Models learned from data are also squishy and vulnerable. As with the octopus, the best strategy is to pursue multiple strategies when it comes to modeling. Even if models are not attacked by sharks, they are still subject to “evolutionary” pressures: If the model sucks, it might eventually be replaced by something better. A machine learning model might be dropped because it lacked causal reasoning. A frequentist model might cause a product to fail if it is used to make predictions even though the generalization error was never measured. An opportunity could be missed because reinforcement learning was never considered.
13.2 The Benefits of Many Mindsets
Knowing many modeling mindsets will make you a more pragmatic and effective modeler.
I adopted new modeling mindsets to turn projects around. This not only helped the projects, but also made me a more pragmatic and effective modeler. By embracing causal inference, my colleague and I were able to improve the (frequentist) statistical model and gain insights into the treatment of axial spondyloarthritis. Switching to supervised learning helped me build better predictive models. Understanding Bayesianism and likelihoodism helped me better understand the frequentist interpretation of probability and recognize its limitations. Embracing supervised learning has helped me built models that are actually decent at prediction. Unsupervised learning opened my perspective about modeling and be more open-minded when starting new projects, doing more exploration than I would do without it. Deep learning unlocked new data types for me (like a side-project to classify x-ray images) and unlocked an end-to-end mindset. And reinforcement learning … this is tougher nut,, and I haven’t done any real projects with it yet, just online courses and lots of reading. And yet, reinforcement learning has shown me how terribly static most other mindsets are, and that models can influence the environment – even if they are not modeled as such.14
I’ve seen experts struggle because they were so narrow-minded in the their approach to modeling. A statistics professors asking strange questions during a machine learning talks, just because he couldn’t fathom that a low generalization error might be a way to evaluate models. If you cling to the modeling mindset you are expert in, even when you experience a limitation, you have trapped yourself within the boundaries of the mindset. These boundaries can prohibit you from becoming and effective modeler. Don’t get me wrong: you can make a career out of becoming the expert in a particular mindset. Especially in academia. As long as you stay in that field, you can excel, like a karateka who has practiced the same kicks, punches and katas thousands of times. But if you put these experts outside their area of expertise, they will get a bloody nose. It’s like sending a karateka into an MMA (mixed martial arts) fight. I would put my money on the well-rounded MMA fighter, not the karateka. So if you want to become an extreme expert in one modeling mindset, make sure you don’t get lured into an street-style modeling task. In a more applied environment, such as working as a data scientist in a company, you don’t have the luxury of mindset purity. Your job is to solve a modeling problem. The more diverse your toolbox, the better your problem-solving abilities. Don’t be the person with just the screwdriver.
The world is messy and doesn’t like to be confined by the limitations of a singular mindset. Be like the octopus and adapt your strategies. Look for non-obvious solutions from other mindsets if you get stuck. Curiosity over pride.
13.3 You Can’t Learn Them All
It’s hard to keep up with new research. Whether it’s causal inference or reinforcement learning, new methods are published daily. There are hundreds of books, online courses and blog posts for each mindset. Even if you stick to established methods, it can take years to gain a deep understanding of a modeling mindset. Realistically, most people focus on one or maybe two or three mindsets. You might start with frequentist statistics, but quickly dive deeper into Bayesian statistics. Or you start with deep learning, but you also pick up supervised and unsupervised learning. As you progress in your modeling journey, you absorb some ideas from other mindsets here and there. You might watch a video about reinforcement learning. You read a paper that uses causal inference. You take a class on hypothesis testing (frequentist inference). Yet it remains unfeasible to master all mindsets.
What’s the solution?
13.4 T-Shaped Modeler
Becoming a T-shaped modeler is the only pragmatic option to knowing many mindsets. The T-shaped modeler embodies a trade-off between being an expert, but also have some general skills in other mindsets. The “T” symbolizes both the breadth and the depth of your modeling skills. A T is flat for most of the horizontal part. But in the middle, it’s deep. This vertical part symbolizes the (few) mindsets that you know in depth. In my case, that would be frequentist statistics and supervised learning. The horizontal part of the T stands for the mindset that you know a little bit. My recommendation: Be excellent in a few mindsets, have a working knowledge in the others.
Becoming excellent in a few mindsets is “easy”. Of course it’s hard work and takes time, but it’s a path that many others have taken before you. The path is clear: language, assumptions and ideas remain the same no matter how deep you wander into the mindset.
The difficult part is getting comfortable with the other mindsets. Of course, you can take a course on the new mindset. But it can be hard to learn it since language, assumptions and ideas are in conflict with the mindsets that you already know well. It’s like leaving an established path, and now you having to carve your own path through the jungle with a machete. Expanding your modeling knowledge requires stepping out of your comfort and challenge deeply held assumptions that you’ve spent years, maybe decades, building.
How does the T-shape modeler work? You work on your model in the way you already know deeply. You work in the vertical part of the T. Then you hit a roadblock: your modeling approach can’t solve the task at hand, because you have reached the vertical end of the T. Each mindset has limitations. An “I”-shaped modeler can’t see the limitations. They will run against the wall and claim that the task is the problem: it can’t be solved. Which is true, if the limitations of the current mindset are accepted. A T-shaped modeler, in contrast, knows when the limitations become problematic for an application. A T-shaped modeler looks left and right for answers instead of repeatedly running their head into the wall.
13.5 After Reading This Book
You have reached the end of this book, and I hope you enjoyed the journey. I leave you with some ideas of what you can do after reading this book:
- If you want to become an effective modeler, you need to create space and time for learning. Set apart a fixed time in the week, or some routine that allows you to dive into new modeling mindsets. Maybe you have an employer who values continuing education and enables you to learn during work hours?
- Pick the modeling mindset that appeals to you the most. Buy a book, read blog posts, take an online course, …
- Talk to people with different mindsets: If you are a machine learner, consider attending a statistics conference; If you are a statistician, engage with the Kaggle community; Attend a deep learning meetup; Sneak into the party of the data mining department; Attend a Bayesian lecture. Join an online data scientist community. Follow causal inference influencers on Twitter. And then just observe and soak up their mindset.
- Revisit Modeling Mindsets whenever you are stuck with a model or before starting a new modeling project.
- Put the Modeling Mindsets book as a physical reminder in a place where you ponder problems: on your desk, as toilet reading, or in your data science book shelf.
Think of predicting traffic jams and showing the predictions to drivers. If drivers now avoid these streets, the predicted traffic jam might never occur: the prediction has changed the outcome, because the model interacts with the environment.↩︎