Researcher Pedro A. Ortega proposes bounded rationality as a unifying framework for decision-making and learning that models samples changing output distributions
Figures compare loss and mutual information across neural net and regression setups.
Generalization is often framed as limiting how much training data can influence the model, using priors, stability, information, or complexity measures. This paper starts from another observation: dependence on data is also where robustness is encoded. It's encoded in the model.
Pedro A. Ortega@AdaptiveAgentsNew paper: Bounded rationality offers a way to unify decision-making and learning. A learner is a decision maker whose sample changes its distribution over outputs. The same response law that improves fit also creates exposure to distortions.
In bounded rationality, decisions trade loss against the cost of reshaping behavior. That cost has a dual role: it prices the movement of the response law, and it determines the perturbations the decision maker is hedged against.
Pedro A. Ortega@AdaptiveAgentsGeneralization is often framed as limiting how much training data can influence the model, using priors, stability, information, or complexity measures. This paper starts from another observation: dependence on data is also where robustness is encoded. It's encoded in the model.