In his article “The Neural Organ Explains the Mind”, Jakob Hohwy (this collection) proposes that the brain, as every other organ in the human body, serves one basic function. Just as one might say that the basic function of the heart is to pump blood through the body or that of the lungs is to provide oxygen, the basic function of the brain is to minimise free energy (Friston 2010). However, this is a very general claim that does not yet establish how the minimisation of free energy is realised in humans. How is this done?
Very generally, the brain stores statistical regularities from the outer environment or, in other words, it forms an internal model about the causal structure of the world. This model is then used to predict the next sensory input. Consequently, we have two values that can be compared with each other: the predicted sensory feedback and the actual sensory feedback. When perceiving, the brain predicts what its own next state will be. Depending on the accuracy of the prediction, a divergence will be present between the predicted and the actual sensory feedback. This divergence is measured in terms of prediction errors. The larger the amount of prediction error, the less accurately the model fits the actual sensory feedback and thus the causal structure of the world. Crucially, the model that fits best, i.e., that which brings forth the smallest amount of prediction error, also determines consciousness. In this framework, free energy amounts to the sum of prediction errors. Thus, minimizing prediction errors always entails the minimisation of free energy.
The minimization of prediction error can generally be achieved in two ways: either the brain can change its models according to the sensory input or, vice versa, it can change the sensory input according to its models. In this scheme the former mode can be seen as veridical perception, whereas the latter can be seen as action, or more formally active inference—the fulfillment of predictions via classic reflex arcs (Friston et al. 2009; Friston et al. 2011). Furthermore, two other factors play a large role in the minimization of prediction error: first, the precision, or “second-order statistics” (Hesselmann et al. 2012), which ultimately encodes how “trustworthy” the actual sensory input is. Precision is realised by synaptic gain, and it has been established that the modulation of precision corresponds to attention (Hohwy 2012). Second, model optimization ensures that models are reduced in complexity in order to account for the largest number of possible states in the long run, i.e., under expected levels of fluctuating noise. For example, sleep has been associated with this type of model optimization (Hobson & Friston 2012). More detailed descriptions of these four factors, i.e., perception, active inference, precision, and model optimization can be found in Hohwy’s article.
Additionally, models are arranged in a cortical hierarchy (Mumford 1992). This hierarchy is characterised, as Hohwy points out (this collection, p. 7), by time and space: models higher up in the hierarchy have a larger temporal scale and involve larger receptive fields than models lower down in the hierarchy, which concern predictions at fast time scales and involve small receptive fields (p. 7). This hierarchy implies a constant message-passing amongst different levels. Once a sensory signal arrives at the lowest level it is compared to the predictions coming from the next higher level (in this case level two).[1] If prediction errors ensue they are sent to the higher level (still level two). Here they are predicted by the next higher level (now level three). This process goes on until prediction errors are minimised to expected levels of noise.
Now the general scheme of prediction error minimization can be presented: the brain builds models that represent the causal structure of the world. These models are, in turn, used to generate predictions about what the next sensory input might be. The two resulting values, i.e., the predicted and the actual sensory feedback, are continuously compared. The divergence between these two values is the prediction error, or free energy. Since it is the brain’s main function to minimise the amount of free energy and therefore prediction error, it will either change its models or engage in active inference. Decisions about which path will be taken depend on the precision of the incoming sensory signal (or prediction error). Signals with high precision are taken to be “trustworthy”, and therefore model changes can follow. Low precision signals, however, require further investigation since noise could be the principal factor in an ambiguous input. In addition, models during wakefulness are changed “on-the-fly”, thus leading to highly idiosyncratic and complex models. This complexity is reduced, for example during sleep (Hobson & Friston 2012), to increase the generalizability of models, since noise is always present.