Learning to Learn
Every machine-learning system you have met was handed its learning procedure by a human: use gradient descent, this learning rate, this schedule. Learning to learn asks why that should be a design decision at all, and treats the procedure itself as something to be learned from experience of many learning problems.
Two loops instead of one
The idea is structural. An inner loop learns one specific task in the ordinary way; an outer loop observes many tasks and improves whatever the inner loop depends on. What the outer loop is allowed to change is what distinguishes the methods:
- The starting point. MAML and Reptile learn an initialisation from which any task in the distribution is a few gradient steps away.
- The optimiser. A recurrent network can be trained to emit the parameter updates itself, replacing hand-written rules such as Adam — the line of work that trained an LSTM to act as the optimiser.
- The representation, or the algorithm wholesale. Metric-learning methods learn an embedding in which unseen classes separate on sight; black-box meta-learners read a task's examples and output predictions with no explicit training step.
The foundation under AutoML
Any AutoML system that gets faster because it has seen other datasets is doing this. Meta-features paired with a portfolio of configurations that worked on similar data — the warm start in Auto-sklearn, the precomputed portfolios in Auto-sklearn 2.0, surrogate models shared across datasets — amount to a learned prior over how to search, not a fixed algorithm. The caveat is implied by the framing: a learned learner transfers only to tasks resembling the distribution it was meta-trained on, which makes the choice of task distribution the real design decision.
Learning to learn, also called meta-learning, is the practice of learning the learning procedure itself from experience across many tasks. An inner loop solves one task; an outer loop improves what that inner loop depends on — the initialisation, the optimiser or the representation — so later tasks need fewer examples and fewer steps. It matters as the principle behind AutoML warm starts: a system that has seen many datasets can narrow the search on an unseen one.
A teacher who has taught a hundred students does not simply know the subject. She knows where people get stuck and which explanation to try first. That second kind of knowledge is about learning itself.
Learning to learn is the idea of giving computers that advantage. An ordinary program is told exactly how to learn: look at the data, adjust, repeat. Here the computer practises on many different learning problems and works out its own better way of learning. The payoff is speed. Faced with something new, it already has a good hunch about where to start, instead of beginning from nothing every single time.