The comments I have provided so far appear to pull somewhat in different directions. I have argued that there is no sharp delineation between functional and mechanistic accounts, and yet I acknowledged that the functional aspects of FEP do set it apart from fully mechanistic accounts. I have argued that merely naming realisers is not explanatory, yet I have acknowledged that mechanistic accounts are explanatory. I have argued (with Harkness) that FEP explains by guiding particular mechanistic accounts, but also by unification. In each of these cases, there seems to be much diversity, or even tension, in how FEP is said to be explanatory.
This diversity and tension, however, is by design. Explanation is not a one-dimensional affair; rather, a hypothesis, h, can be explanatory in a number of different ways. This can be seen by applying the overall Bayesian framework to scientific explanation itself. The strength of the case for h is consummate with how much of the evidence, e, h can explain away. As we know from the discussion of FEP, explaining away can happen in diverse ways: by changing the accuracy, the precision, or the complexity of h, or by intervening to obtain expected, high precision e. As discussed for FEP in Hohwy (this collection), we can also consider h’s ability to explain away e over shorter or longer time scales: if h has much fine-grained detail it will be able to explain away much of the short term variability in e but may not be useful in the longer term, whereas a more abstract h is unable to deal with fine-grained detail but can better accommodate longer prediction horizons.
Sometimes these diverse aspects of Bayesian explaining-away pull in different directions. For example, an attempt at unification via de-complexifying h may come at the loss of explaining some particular mechanistic instantiations. Conversely, an overly complex h may be overfitted and thereby explain away occurrent particular detail extremely well but be at a loss in terms of explaining many other parts of e.
In constructing a scientific explanation, how should one balance these different aspects of Bayesian explanation? Again we can appeal to FEP itself for inspiration: a good explanation minimizes prediction error on average and in the long run. That is, a good explanation should not generate excessively large prediction errors, and should be robust enough to persist successfully for a long time. This is intuitive, since we don’t trust explanations that tend to generate large prediction errors, nor explanations that cease to apply once circumstances change slightly.
Formulating the goal of scientific explanation in this way immediately raises the question of what it means for prediction error to be “large” or for a hypothesis to survive a “long time”. The answer lies in expected precisions and context dependence. In building a theory, the scientist also needs to build up expectations for the precision (i.e., size) of prediction errors, and for the spatiotemporal structure of the phenomenon of interest. Not surprisingly, these aspects are also found in the conception of hierarchical Bayesian inference.
Achieving this balanced goal requires a golden-mean-type strategy: explanations should not be excessively general nor excessively particular, given context and expectations. That is, h should be able to explain away e in the long term without generating excessive prediction errors in the short term, as guided by expectations of precision and domain.
I think FEP is useful for attaining this golden mean, and that this is what makes FEP so attractive and promising. As a scientific hypothesis, it does not prioritise one type of explanatory aspect over another, but instead balances explanatory aspects against each other such that prediction error concerning the workings of the mind is very satisfyingly minimized on average and in the long run (and this indeed is the message of The Predictive Mind). Rather poetically, in my view, this means that we should evaluate FEP’s explanatory prowess by applying it to itself.