Editing Meta Learning (section)

== <span style="color: #FFFFFF;">Understanding</span> ==
Standard training gives a model a fixed behavior. Meta-learning gives a model the ability to '''quickly adapt''' its behavior given a few new examples.

'''The meta-learning objective''': Across many tasks T sampled from a distribution p(T), find model parameters θ that can quickly adapt to any task using only a few examples. Formally: min''θ E''{T~p(T)}[L''T(f''{θ'})] where θ' = Adapt(θ, support''set''T).

'''Three meta-learning approaches''':

'''Metric-based''': Learn an embedding space where classification is easy — similar examples are close, different ones are far. At test time, classify by distance to class prototypes (Prototypical Networks) or by weighted attention over support examples (Matching Networks).

'''Optimization-based (MAML)''': Find model initialization θ such that a few gradient steps on the support set produce a good model for the query set. The meta-update optimizes through the adaptation process — it literally backpropagates through gradient descent steps.

'''Model-based''': Use a recurrent or attention architecture that quickly updates its "memory" when shown support examples. The model's hidden state encodes the task context, enabling immediate adaptation.

'''In-context learning''' (emergent in LLMs) is meta-learning without gradient updates: GPT-4 can learn to translate into a new language, write in a new style, or follow new formatting rules from just a few examples in the prompt. The model's weights don't change — it "adapts" purely through the attention mechanism reading the context.
</div>

<div style="background-color: #8B0000; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">