Active inference has appeared repeatedly as an important concept throughout this paper. Yet it is more difficult to grasp than the basics of PP, which involve passive predictive inference. This is partly because several senses of active inference can be distinguished, which have not previously been fully elaborated.
In general, active inference can be harnessed to drive action, or to improve perceptual predictions. In the former case, actions emerge from the minimization of proprioceptive prediction errors through engaging classical reflex arcs (Friston et al. 2010). This implies the existence of generative models that predict time-varying flows of proprioceptive inputs (rather than just end-points), and also the transient reduction of expected precision of proprioceptive prediction errors, corresponding to sensory attenuation (Brown et al. 2013).
In the latter case, actions are engaged in order to generate new sensory samples, with the aim of minimizing uncertainty in perceptual predictions. This can be achieved in several different ways, as is apparent by analogy with experimental design in scientific hypothesis testing. Actions can be selected that (i) are expected to confirm current perceptual hypotheses (Friston et al. 2012); (ii) are expected to disconfirm such hypotheses; or (iii) are expected to disambiguate between competing hypotheses (Bongard et al. 2006). A scientist may perform different experiments when attempting to find evidence against a current hypothesis than when trying to decide between different hypotheses. In just the same way, active inference may prescribe different sampling actions for these different objectives.
These distinctions underline that active inference implies counterfactual PP. In order for a brain to select those actions most likely to confirm, disconfirm, or decide between current predictive model(s), it is necessary to encode expected sensory inputs and precisions related to potential (but not executed) actions. This is evident in the example of oculomotor control described earlier (Friston et al. 2012). Here, saccades are guided on the basis of the expected precision of sensory prediction errors so as to minimize the uncertainty in current perceptual predictions. Note that this study retained the higher-order prior that only a single perceptual prediction exists at any one time, precluding active inference in its disambiguatory sense.
Several related ideas arise in connection with these new readings of active inference. Seeking disconfirmatory or disruptive evidence is closely related to maximizing Bayesian surprise (Itti & Baldi 2009). This also reminds us that the best statistical models are usually those that successfully account for the most variance with the fewest degrees of freedom (model parameters), not just those that result in low residual error per se. In addition, disambiguating competing hypotheses moves from Bayesian model selection and optimization to model comparison, where arbitration among competing models is mediated by trade-offs between accuracy and model complexity (Rosa et al. 2012).
The information-seeking (or “infotropic”[15]) role of active inference puts a different gloss on the free energy principle, which had been interpreted simply as minimization of prediction error. Rather, now the idea is that systems best ensure their long-run survival by inducing the most predictive model of the causes of sensory signals, and this requires disruptive and/or disambiguating active inference, in order to always put the current-best model to the test. This view helps dissolve worries about the so-called “dark room problem” (Friston et al. 2012), in which prediction error is minimized by predicting something simple (e.g., the absence of visual input) and then trivially confirming this prediction (e.g., by closing one’s eyes).[16] Previous responses to this challenge have appealed to the idea of higher-order priors that are incompatible with trivial minimization of lower-level prediction errors: closing one’s eyes (or staying put in a dark room) is not expected to lead to homeostatic integrity on average and over time (Friston et al. 2012; Hohwy 2013). It is perhaps more elegant to consider that disruptive and disambiguatory active inferences imply exploratory sampling actions, independent of any higher-order priors about the dynamics of sensory signals per se. Further work is needed to see how cost functions reflecting infotropic active inference can be explicitly incorporated into PP and the free energy principle.
What can be said about counterfactual PP and active inference when applied to interoception? Is there a sense in which predictive models underlying emotion and mood encode counterfactual associations linking fictive interoceptive signals (and their likely causes) to autonomic or allostatic controls? And if so, what phenomenological dimensions of affective experience depend on these associations? While these remain open questions, we can at least sketch the territory.
We have seen that active inference in exteroception implies counterfactual processing, so that actions can be chosen according to their predicted effects in terms of (dis)confirming or disambiguating sensory predictions. The same argument applies to interoception. For active interoceptive inference to effectively disambiguate predictive models, or (dis)confirm interoceptive predictions, predictive models must be equipped with counterfactual associations relating to the likely effects of autonomic or (at higher hierarchical levels) allostatic controls. At least in this sense, interoceptive inference then also involves counterfactual expectations.
That said, there are likely to be substantial differences in how counterfactual active inference plays out in interoceptive settings. For instance, it may not be adaptive (in the long run) for organisms to continually attempt to disconfirm current interoceptive predictions, assuming these are compatible with homeostatic integrity. To put it colloquially, we do not want to drive our essential variables continually close to viability limits, just to check whether they are always capable of returning. This recalls our earlier point (section 4.1) that predictive control is more naturally applicable to interoception than exteroception, given the imperative of maintaining the homeostasis of essential variables. In addition, the causal structure of counterfactual associations encoded by interoceptive predictive models is undoubtedly very different than in cases like vision. These differences may speak to the substantial phenomenological differences in the kind of perceptual presence associated with these distinct conscious contents (Seth et al. 2011).