There is a different worry about preposterousness, also related to the issue of evidence. This worry is that the free energy principle is so general that anything the brain does can be construed as minimizing its prediction error. This is most clearly seen once the idea is cast in Bayesian terms. The brain harbours priors about the causes in the environment, and it calculates likelihoods that it combines with the priors to arrive at posterior probabilities for the hypotheses in question. One way to make this story apply to a particular case is to ascertain what is believed and then in a retrodictive fashion, posit priors and likelihoods accordingly unto the brain in question. If this can always be done, then the theory is trivialised by “just-so” stories and explains nothing. It is then preposterous because it pretends to be fundamental but is just trivial.
This triviality worry alerts the defender of the free energy principle to some pitfalls, but it is not a critical worry. To see this, an appeal can again be made to the theory of evolution. It is clear that when described in very general terms, anything can be described as enhancing fitness. For example, in an infamous hoax, Ramachandran gave a ridiculous, just-so adaptationist account of why gentlemen prefer blondes (Ramachandran & Blakeslee 1998), which some reportedly took seriously. Yet, no one serious thinks this invalidates the theory of evolution. The reason is, to repeat, that there is abundant solid, non-trivial evidence in favour of evolution. In other words, the presence of just-so triviality at some level of description can co-exist with non-trivial explanations at the level of detailed, quantifiable evidence. Therefore the free energy principle cannot be invalidated just because it invites just-so stories. Of course, it is then hostage to translation into more precise, constricted applications to various domains, where predictions can be quantified and just-so stories avoided. Though there is nowhere near the same evidence that we have for the theory of evolution, evidence of this sort is becoming available (some is reviewed in Hohwy 2013).
Whereas the triviality worry does not invalidate the free energy principle, it does alert to some pitfalls. In particular, when forming hypotheses and when explaining phenomena in Bayesian terms, priors should not be stipulated independently of other evidence. If there is independent reason for asserting a prior with an explanatory role, then it is less likely that this prior is part of a just-so story. Similarly, discovery and manipulation of priors has a particularly important role in the defence of the free energy principle as applied to perception. For example, there is independent evidence that we expect light to come more or less from above (Adams et al. 2004), that objects move fairly slowly (Sotiropoulos et al. 2011), and that we expect others to look at us (Mareschal et al. 2013). Once established on independent grounds, researchers are better able to appeal to such priors in other explanations. This then helps avoid the just-so pitfall.
The triviality worry was that everything we do can be made to fit with the free energy principle. A different worry is that almost nothing we do fits with the free energy principle. If the free energy principle basically says the brain is an organ that tries to slow down the causal impact upon it from the world, then why don’t organisms with brains just seek out sensory deprivation such as dark, silent rooms (Friston et al. 2012)? This dark room problem is aired very often and is natural on first thought when considering prediction error minimization. However, it also rests on a fundamental misreading of the free energy principle. The principle is essentially about maintaining the organism in its expected states, homeostatically defined, on average and in the long run. Locking oneself up in a dark silent room will only produce transitory free-energy minimization, as the demands of the world and the body will not be avoided for long. Soon, action is required to seek food, and soon the local council will come round to switch off the gas. It is much better for the brain to harness the deep model of the world in order to control its movement through the environment and thereby maintain itself more efficiently in its expected states.
Notice that this point harks back to a very basic hyperprior mentioned above—namely that the world is a changing place so that occupying the same state for too long will incur increasing free energy costs. This means that even if you currently have the prior that sensory deprivation is the right strategy for minimizing free energy, and even if this strategy works initially (as it does after a long and stressful day), that prior will decrease in strength as time goes by—leading to action and thus escape from sensory deprivation.
This response to the dark room problem in fact has a parallel in evolutionary theory. It has been argued that the free energy principle is false, essentially because not every action contributes directly to instantaneous predictionerror minimization and, analogously, it could be objected that evolutionary theory is false because not every trait directly contributes to instantaneous fitness. But of course this is a poor objection because fitness is measured over longer timescales and some traits, such as spandrels, contribute indirectly to fitness.
This and the preceding two sections have considered whether the explanatory ambition of the free energy principle is preposterous. By comparing the principle with the theory of evolution, and casting the worry in terms of philosophy of science, it can be seen that the explanatory ambition is not preposterous in and of itself. The verdict on the principle must come down to the quality of the explanations it offers and the amount of evidence in its favour. The principle is bound to be controversial, however, because it strongly explains away competing theories.
Of course, there are further issues to explore regarding the analogy between the free energy principle and the theory of evolution, and no doubt the analogy will have its limits. One interesting issue concerns the possibility of theory revision and thereby the possibility that the original statement of a theory is strictly speaking, false, even if it is one of those theories with extreme explanatory scope. The notion of natural selection as the only mechanism behind evolution is, for example, put under pressure by the discovery of genetic drift. This has led to revision of the theory of evolution, to encompass drift. Could something similar happen to the free energy principle, or is it in effect so ambitious that it is unrevisable? Conversely, is there any conceivable evidence that could falsify the current version of the theory in a wholesale fashion, rather than the piecemeal, detailed fashion discussed above? There are various answers available here, all of which reflect the peculiar theory emerging from the free energy principle.
First, the current form of the principle itself results from a long series of revisions of the basic idea that the brain engages in some kind of inference. Helmholtz’ and Ibn Al Haytham’s original ideas (reviewed briefly in Hohwy 2013) have been greatly revised, particularly in response to the mathematical realisation that the inversion of generative models presents an intractable problem, thus calling for variational Bayesian approaches to approximate inference. These developments occured partly in concert with the empirical discovery that the brain, as mentioned above, is characterized by massive backwards connectivity. It is then not unreasonable to say that older feed-forward versions of computational, information theoretical (e.g., infomax) theories of cognition constitute earlier versions of the free energy principle and that the latter is a revision in the light of formal and empirical discovery. The analogy with theory of evolution can thus be maintained in at least this backwards-looking respect.
Second, a more forward-looking example concerns the nature of the backwards connectivity in the brain. The free energy principle deems these descending signals predictions, but crucially it needs them to be of two kinds, namely predictions of the means of the underlying level’s representations, and, as mentioned briefly above, predictions of the precisions of the underlying representations (thus encompassing sufficient statistics). There is some direct and some circumstantial evidence in favour of this dual role for descending signals, but the empirical jury is still out. Should it be found that descending signals do not mediate expected precisions, this would falsify the free energy principle. Notice that this falsification would be specific to the free energy principle, since the element of expected precisions is not found in some of the much broader theories in the academic marketplace that seem to countenance a predictive element in cognition. Notice also that a failure to identify top-down expectations of precision would amount to a wholesale falsification of the principle, since these “second-order” expectations are crucial not only for perception but also for action and action initiation (as explained above).
Third, and speaking much more generally, the principle would be falsified if a creature was found that did not act at all to maintain itself in a limited set of states (in our changing world). Such a creature should not on average and over time change its model parameters or active states and yet it would be able to prevent itself from being dispersed with equal probability among all possible states. This is a clear notion of a strong falsifier, and it speaks to the beauty of the free energy principle since it showcases its deep link between life and mind. However, it is not a very feasible falsifier because there is significant doubt that we would classify such a ‘creature’ as being alive or being a creature at all. Consider, for example, that a simple rock would serve as a falsifier in this sense since it is maintained on average and over the long run (that is, its states do not immediately disperse). One possibility here (Friston 2013) is to require that the scope is restricted to creatures that are space-filling, that is, who visits the individual states making up their overall set of expected states. A falsifier would then be a creature that manages to be space-filling but who does not manage this by changing its internal and active states via variational Bayes.
One nice question, in all of this, is whether the theory of evolution and the free energy principle can co-exist—and if so, how. This is a substantial issue, and a pertinent one, since both theories are fundamental and pertain to some of the same aspects—such as morphology, phenotypes, and life. Here is not the place to try to answer this interesting question, though inevitably some initial moves are made that might start to integrate them.