In the Body’s Eye: The computational anatomy of interoceptive inference
Fig 6
Computational neuroanatomy of interoception.
The schematic above shows the form of the neuronal message passing implied by active inference for the generative model depicted in Fig 2. We have related this to the anatomical networks that could implement these inferences. The sensory observations in our simulations are visual and interoceptive (cardiac). These sensations are carried by cranial nerves II and IX respectively. Cranial nerve II targets the superior colliculus in the midbrain. This structure sends short latency visual data to the amygdala, which is well placed to make inferences about emotionally salient stimuli. The amygdala additionally receives downstream visual data from the ventral visual stream in the temporal lobe. Cranial nerve IX carries information from the carotid sinus baroreceptors to the nucleus tractus solitarus in the brainstem. This nucleus communicates with the posterior insula (via thalamic and PAG relays); the anterior cingulate monitors and controls the precision of this ascending visceral information via neuromodulation, possibly via feedback through noradrenergic pathways (not shown). The posterior insula and amygdala interact with one another but also project to the anterior insula. This targets the nucleus ambiguus (via brainstem relays such as the periaqueductal gray), which gives rise to the vagus (X) nerve. The vagus nerve targets neurons in the cardiac plexus that project to both the sinoatrial node and the atrioventricular node of the heart, slowing its rhythm. The nucleus tractus solitarus additionally participates in a reflex loop implicating the sympathetic control of the cardiac cycle, but this is omitted for simplicity. The functional anatomy suggested here implies the anterior insula might play a similar computational role in autonomic policy selection to the basal ganglia in selection of policies involving the skeletal muscles [114]. Note that inscribed directed influences (blue arrows), are not assumed to be monosynaptic–for simplicity, many intermediary relay nodes have been omitted.