摘要
Motivation to engage in any epistemic behavior can be decomposed into two basic types that emerge in various guises across different disciplines and areas of study. The first basic dimension refers to a desire to approach versus avoid nonspecific certainty, which has epistemic value. It describes a need for an unambiguous, precise answer to a question, regardless of that answer’s specific content. Second basic dimension refers to a desire to approach versus avoid specific certainty, which has instrumental value. It concerns a need for the specific content of one’s beliefs and prior preferences. Together, they explain diverse epistemic behaviors, such as seeking, avoiding, and biasing new information and revising and updating, versus protecting, one’s beliefs, when confronted with new evidence. The relative strength of these motivational components determines the form of (Bayes optimal) epistemic behavior that follows. People often seek new information and eagerly update their beliefs. Other times they avoid information or resist revising their beliefs. What explains those different reactions? Answers to this question often frame information processing as a competition between cognition and motivation. Here, we dissolve this dichotomy by bringing together two theoretical frameworks: epistemic motivation and active inference. Despite evolving from different intellectual traditions, both frameworks attest to the indispensability of motivational considerations to the epistemic process. The imperatives that guide model construction under the epistemic motivation framework can be mapped onto key constructs in active inference. Drawing these connections offers a way of articulating social psychological constructs in terms of Bayesian computations and provides a generative testing ground for future work. People often seek new information and eagerly update their beliefs. Other times they avoid information or resist revising their beliefs. What explains those different reactions? Answers to this question often frame information processing as a competition between cognition and motivation. Here, we dissolve this dichotomy by bringing together two theoretical frameworks: epistemic motivation and active inference. Despite evolving from different intellectual traditions, both frameworks attest to the indispensability of motivational considerations to the epistemic process. The imperatives that guide model construction under the epistemic motivation framework can be mapped onto key constructs in active inference. Drawing these connections offers a way of articulating social psychological constructs in terms of Bayesian computations and provides a generative testing ground for future work. the expected log likelihood of some outcome, under posterior Bayesian beliefs about the causes of that outcome. the minimization of variational free energy through approximate Bayesian inference and active sampling of (sensory) data. This sampling induces belief updating, under prior beliefs that sampling will minimize free energy expected in the future. This is equivalent to resolving uncertainty and maximizing model evidence, sometimes called self-evidencing. the expected inaccuracy of future outcomes, under a particular policy, as measured by the conditional entropy (i.e., uncertainty) about outcomes, given their causes. the process of statistical inference, in which Bayes' theorem is used to update the probability of a hypothesis as more evidence or information becomes available. Technically, a prior belief is updated to form a posterior belief. Bayesian belief refers to a posterior probability distribution over a random variable, such as a latent cause or hidden state of the world causing (sensory) data. the divergence between prior and posterior beliefs; in other words, the degree to which Bayesian beliefs change before and after belief updating. the information gain expected under a particular policy. This is sometimes referred to as intrinsic motivation, Bayesian surprise, or salience. Novelty is a form of salience that reflects the epistemic affordance of policies, which resolve uncertainty about the parameters of a generative model. an attribute of a policy that can be decomposed into epistemic and pragmatic value, or into risk and ambiguity (see Box 2 in the main text). a probability distribution over the causes of observable consequences. A generative model is usually specified in terms of a likelihood and a prior belief; namely, the probability of an outcome given its cause and the prior belief about the course. the optimization of beliefs by maximizing model evidence. Approximate Bayesian inference corresponds to minimizing variational free energy. the difference between two probability distributions, as measured with their relative entropy. the probability of some (sensory) data under a generative model; also known as the marginal likelihood. The log model evidence is approximated by (negative) variational free energy. an ordered sequence of actions. a belief about the causes of (sensory) data after belief updating. a belief about the causes of data, prior to sampling (sensory) data. a prior belief about an outcome in the future, which generally depends upon a policy. the complexity expected under a particular policy. In other words, the expected divergence between predicted and preferred outcomes. a functional of sensory data and posterior beliefs that approximates model evidence. Free energy scores the implausibility of some (sensory) data, given posterior beliefs about the causes of those data.