Learning to Learn: A Narrative Review of Meta-Learning from Synaptic Rules to Model-Agnostic Adaptation
Year: 2026
Type: review
Abstract: Meta-learning---the learning whose object is the learning itself---moved from Schmidhuber's self-referential principles and Bengio's synaptic rules through Thrun's learning to learn, Hochreiter's bidirectional LSTM meta-learner, and Ando and Zhang's predictive structures to the few-shot wave of Vinyals's matching networks, Duan's RL squared, Ravi's optimization-as-model, Finn's MAML, and Snell's prototypical networks, and Hospedales's survey. This article presents a narrative review of that arc's canonical line: Schmidhuber's 1987 principles, Bengio, Bengio, and Cloutier's 1991 synaptic rules, Thrun's 1996 n-th thing, Thrun and Pratt's 1998 Learning to Learn, Hochreiter and colleagues's 2001 bidirectional LSTM, Ando and Zhang's 2005 structures, Vinyals and colleagues's 2016 matching networks, Duan and colleagues's 2016 RL squared, Ravi and Larochelle's 2017 optimization as model, Finn, Abbeel, and Levine's 2017 MAML, Snell, Swersky, and Zemel's 2017 prototypical networks, and Hospedales and colleagues's 2021 survey. The review is organized around three themes: the learning-to-learn origins, in which the self-reference's, the synaptic's, and the transfer's framings defined the field's questions; the few-shot wave, in which the metric's, the recurrent's, the gradient's, and the prototype's methods made the fast adaptation practical; and the survey's synthesis, in which the taxonomy's three families organized the field's maturity. It is concluded that meta-learning is the learning's second order---and that its arc is the adaptation's generalization from the task's specifics to the learner's rules.
Authors: Revista, Zen, IA, 10
Institution: Faculdades Guarulhos
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204699720