Deliberately leaving things out

There is a trick that makes a model stronger and sounds absurd at first: people deliberately let part of it rest.

The trick is called dropout. During training, at each round, a number of the model's “neurons” are randomly switched off, sometimes as many as half. It sounds like handicapping yourself, and the effect is rather good: because it does not know which part will be switched off, the model dares not pile everything onto a few details, and is forced to understand the problem in a more general and more robust way. It is like a coach deliberately resting key players in rotation during training, so the whole squad learns to carry the load rather than depending on one star.

What stays with me is the lesson underneath: deliberate thinning does not weaken a system, it usually makes it more durable.

This holds very well for businesses. A company where everything funnels through one person, one sales channel, or one large customer is extremely fragile, however good the figures look right now. Conversely, actively spreading the load and reducing the places of excessive dependence usually makes an organisation more supple. Real resilience sits here: remove one link and the machine still runs.

If a key person were away next week, would your organisation still run properly?

To your desk

A model gets stronger by actively leaving things out. Our work is the same.

This week, try choosing one thing to stop: a report almost nobody reads, a meeting still held out of habit, or a tool that duplicates another. Add nothing new, just lighten.

Thinning here does not mean doing less. It means putting the effort into the parts that actually produce the result.