Precision, synaptic gain and neuromodulation

Friston, Daunizeau and Kiebel (2009) used precision as a parameter in a simulated agent. They set a log-precision of 16, which they said corresponds to a standard deviation of 0.00034, so that predicted sensory states were dominated by the agent's prior expectations¹. The prediction error was then explained away by action rather than by revised predictions¹. They described conditional precisions as quantities that control the relative influence of bottom-up sensory information and top-down predictions, and said these mediate attentional gain in a perceptual setting². They added that, in biological formulations of the free-energy principle, current thinking is that dopamine might encode the precision of prior expectations³.

Friston and colleagues (2012) tied precision to uncertainty due to random fluctuations or noise. By associating salience with precision, they linked it to incentive salience in psychology and aberrant salience in psychopathology⁴. They noted that the idea of dopamine modulating synaptic gain plays a key role in several proposals⁴. To test the account, they repeated their simulations at different precision levels in the putative targets of dopaminergic projections. This was meant to simulate depleted dopamine acting at postsynaptic D1 receptors to reduce postsynaptic gain⁵. Log precision was reduced by 50% in six steps in superior colliculus cells reporting prediction errors on the salience of target locations⁵.

Sengupta, Stemmler and Friston (2013) listed three things that perception optimises to minimise prediction error: synaptic activity, synaptic efficacy and synaptic gain. They assigned gain to attention and salience⁶. Constant and colleagues (2018) used salience in a narrower formal sense. For them it is information gain, reduction of uncertainty, Bayesian surprise or epistemic value, forming part of expected free energy⁸. Cues become salient because they indicate actions that resolve uncertainty⁸. The held texts therefore use the word in two ways: as the precision of a representation⁴ and as epistemic value⁸.

Friston, Levin, Sengupta and Pezzulo (2015) applied the scheme to pattern regulation. In their model, the internal states of each cell organise themselves to minimise precision weighted prediction error⁷. They suggested the encoding of target locations could be associated with a genetic code, and its transcription into predictions with epigenetic processes⁷. Pio-Lopez and colleagues (2022) stated that in active inference the precision of sensory information is assumed to be mediated by neuromodulators such as dopamine or acetylcholine⁹. In their simulation they reduced sensitivity to the signals of other cells in two cells with too high precision. They compared this to D2 antagonist antipsychotics and suggested that dopaminergic antagonists could be tested in morphogenesis⁹.