A concept · Inside the Machine · Computing
Gradient Descent
Improve by always stepping downhill on the error.
A colleague's spreadsheet forecast is wrong every month, and every month she adjusts it a little towards what actually happened. She never has a theory. After a year it is the most accurate forecast in the office. Meanwhile we read that a model trained for months on thousands of chips has learned to write, and we assume something like understanding must have switched on partway through. What if the model was only ever doing what she does?
What Gradient Descent means
Gradient descent is the learning method behind nearly every modern AI: measure how wrong the output was, nudge every internal setting a fraction in the direction that would have made it less wrong, and repeat, millions of times. Nothing else happens. Owning that removes a lot of mystique. We stop picturing a moment of insight inside the machine and start picturing a very long walk downhill on a landscape of error. We start asking the useful questions: what counted as the error, and who chose it, because the model became whatever made that number smaller. Before: it learned to write. After: it was pushed, tiny step by tiny step, towards text that scored well, and whatever scored well is what we get.
Not to be confused with
Gradient descent is not backpropagation, though the two run together. Backpropagation works out which way is downhill for every setting in the network. Gradient descent is the step taken in that direction. One finds the slope; the other walks it.
Gradient Descent examples: where it shows up
- At work: a tool that is "improving" on our feedback is only being pushed towards whatever we rated highly, so the ratings we give carelessly are the ones it learns hardest.
- With money: a trading model tuned to last year's prices walked downhill on last year's errors, and a new year is a different landscape.
- In the news: when a model turns out to have picked up a bias, no one put it there; it was the shortest path to a smaller error on the data it was given.
Where it sits
Concept 2 of 4 in Inside the Machine, a journey in the course Understand AI.
See where Gradient Descent sits on the sky →
Nearest on the sky
A filled dot marks a star that is already open to walk.
Knowing it isn't owning it.
Connect Gradient Descent to your mental models and apply it to your world today.
The first one is on us. No signup, no card.