In his very rich target paper, Anil Seth calls attention to one of the less well-considered precursors of PP: cybernetics. A central concept of cybernetics is the notion of homeostasis, which denotes an equilibrium of the system’s parameters. This equilibrium is maintained by keeping the system’s essential variables, like levels of blood oxygenation or blood sugar (cf. Seth this collection, p. 7), within a certain range (cf. ibid. pp. 7-8.). The process of achieving homeostasis is called allostasis (cf. ibid. p. 8). Cybernetic systems are teleological, i.e., goal-directed, because they are always trying to reach and preserve homeostasis. This suggests that control is more important than perception (cf. ibid. p. 9), and, as Seth emphasizes, it prioritizes interoceptive control over exteroceptive control: the main goal is to control the system’s essential variables; interaction with the world is only necessary to the extent that it affects these variables (ibid. pp. 9-10.).
The principles of cybernetics fit astonishingly well to ideas motivating Karl Friston’s FEP (which can, in some respects, be seen as a generalization of predictive processing).[2] The fundamental assumption behind this principle is that biological systems seek to “maintain their states and form in the face of a constantly changing environment” (Friston 2010, p. 127). This is obviously similar to the goal of achieving homeostasis.[3] Another focus of FEP is active inference, because action can reduce the surprisal of the agent’s states (which is necessary to “resist a tendency to disorder”, Friston 2009, p. 293); perceptual inference can only reduce the free-energy bound on surprise (Friston 2009, p. 294). This is in stark contrast with the Helmholtzian roots of PP, according to which action is primarily in the service of perception:
[...] wir beobachten unter fortdauernder eigener Thätigkeit, und gelangen dadurch zur Kenntniss des Bestehens eines gesetzlichen Verhältnisses zwischen unseren Innervationen und dem Präsentwerden der verschiedenen Eindrücke aus dem Kreise der zeitweiligen Präsentabilien. Jede unserer willkührlichen Bewegungen, durch die wir die Erscheinungsweise der Objecte abändern, ist als ein Experiment zu betrachten, durch welches wir prüfen, ob wir das gesetzliche Verhalten der vorliegenden Erscheinung, d.h. ihr vorausgesetztes Bestehen in bestimmter Raumordnung, richtig aufgefasst haben.[4] (Helmholtz 1959, p. 39)
According to this view, the main target of action is to find confirmatory evidence for internally-generated hypotheses. In short, the contrast between these two views can be described as “action as hypothesis-testing” versus “action as predictive control”. Whereas the first seems to fit best to the Helmholtzian roots of PP (and puts action in the service of perception), the second seems to fit better to its cybernetic origins. Most notably, the free-energy principle combines both aspects, but assigns a pivotal role to action (perceptual inference only makes the free-energy bound on surprise tight, active inference leads to a further reduction of free energy, reducing surprise implicitly).
Seth compares model selection and optimization in evolutionary robotics to how these processes are implemented in active inference (pp. 14-15.). He cites the famous starfish robot developed by Josh Bongard, Victor Zykov, & Hod Lipson (2006) as an example. In a first phase, the robot generates multiple competing models of its own morphology and performs actions for which these models predict different sensory feedback. By comparing these predictions to the actual feedback, the starfish can thus exclude some of the possible models. When the robot has eliminated all but one model, a second phase starts and it uses this model to control its body and generate walking behavior (action as predictive control). Crucially, when the robot’s morphology changes (when an experimenter removes one of its limbs), it can switch back to the first phase, re-creating competing models and using action to eliminate most of them (action as hypothesis-testing).
Seth points out that the second phase, in which the robot walks around, suggests that the main purpose of predictive models is to control behavior effectively, regardless of how accurately it represents the world or the body (p. 15). In the first phase, by contrast, exploratory actions are conducted in order to learn something about the body, not to reach a goal involving its environment (ibid.). As noted above, such instances of action conform more to Helmholtzian than to cybernetic roots (action as hypothesis-testing).
What this shows is that action can fulfill different purposes—not just theoretically, but also in real applications. The robot starfish uses action in at least two ways. Drawing on the often-noted analogy between PP and scientific practice (cf. Gregory 1980), Seth explores further purposes of action. This leads to a distinction between three types of active inference (pp. 18f.). The first involves active sampling to confirm predictions derived from currently active models; the second is employed to seek evidence that would disconfirm currently held hypotheses; the third involves sampling in order to disambiguate between alternative hypotheses (p. 19).
Crucially, Seth does not elaborate much on the notion of falsification or disconfirmation. He relates disconfirmation to Bayesian surprise (which formalizes the extent to which new evidence leads to a revision of prior representations, cf. Baldi & Itti 2010). Accordingly, he characterizes seeking falsifying evidence in terms of maximizing Bayesian surprise. However, the paper quoted in this context, Itti & Baldi (2009) only investigates the hypothesis that surprising information attracts attention, not that subjects act to maximize surprise. Friston et al. (2012, p. 6) clarify the relation between FEP and maximization of Bayesian surprise:
The term Bayesian surprise can be a bit confusing because minimizing surprise per se (or maximizing model evidence) involves keeping Bayesian surprise (complexity) as small as possible. This paradox can be resolved here by noting that agents expect Bayesian surprise to be maximized and then acting to minimize their surprise, given what they expect.
In the following section, I will clarify the notion of falsification, and discuss the ways in which it is used in PP. More specifically, I will illustrate various types of active inference by drawing a slightly broader analogy with theory of science. In particular, I will consider views put forward by Karl Popper and Thomas Kuhn, respectively. This will serve to help us get a handle on the general merits of confirmation and disconfirmation. Furthermore, both Popper’s falsificationism and Kuhn’s paradigm change can be related to aspects of predictive processing, which will hopefully lead to a better understanding of hypothesis-testing in PP. As a consequence, I invite Seth to provide a refined treatment of the relation between falsification and active inference.