The anatomy of the cochlea lends itself remarkably well to interpretation through the Free Energy Principle. This simulation explores that correspondence — mapping cochlear structures onto the architecture of Active Inference.
Any system persisting at a non-equilibrium steady state possesses a Markov blanket that separates internal from external states. The system minimises variational free energy:
The cochlea can be read as a Markov blanket in physical form. The basilar membrane and Reissner membrane separate endolymph (internal, +80mV, high K⁺) from perilymph (external, ~0mV, high Na⁺). The stria vascularis maintains this electrochemical partition as the metabolic engine.
One row of inner hair cells (IHCs) serves as the sensory surface. Stereocilia deflection is bistable: toward the tallest row triggers depolarisation, away triggers hyperpolarisation. In Active Inference terms, each mechanotransduction event generates one unit of sensory evidence.
95% of afferent nerve fibres synapse on IHCs, each type I fibre contacting a single IHC. This anatomical specificity is consistent with high-precision sensory encoding.
Three rows of outer hair cells (OHCs) form the cochlear amplifier. The motor protein prestin (Liberman et al., 2002) enables electromotility: voltage-driven length changes that amplify basilar membrane vibration by 40–60 dB (Dallos, 2008).
The OHC amplifier can be interpreted as implementing prior precision control: it amplifies anticipated frequencies before they reach the sensory surface. The medial olivocochlear (MOC) efferent system modulates OHC gain from the brainstem (Guinan, 2006), providing descending precision control — consistent with the Active Inference framework where top-down predictions modulate sensory gain.
Sound enters at the cochlear base (high frequency, narrow basilar membrane) and propagates as a traveling wave toward the apex (low frequency, wide membrane). At each location, the wave accumulates energy that matches the local resonant frequency.
In the language of Active Inference, this can be read as evidence accumulation: sensory data builds toward a belief update at the characteristic frequency location, after which the signal decays — analogous to a prediction error being resolved.
The pillar cells of the tunnel of Corti physically separate IHC from OHC rows. Nerve fibres cross through this boundary (habenula perforata). In Markov blanket terms, this anatomical partition separates the sensory channel from the amplification channel — a physical boundary between evidence and prior.
The cochlear tuning curve exhibits a sharp excitatory tip at the characteristic frequency, surrounded by inhibitory sidebands via two-tone suppression. This centre-surround structure parallels the receptive field organisation seen throughout the sensory nervous system.
The balance between sensory precision (IHC) and prior precision (OHC amplifier) determines cochlear processing. When OHC gain is high, the system amplifies anticipated frequencies — exploiting prior expectations. When gain is low, the raw sensory input dominates.
When OHCs are damaged, the cochlear amplifier fails. The brain increases central gain to compensate for reduced sensory input. In predictive coding terms, this is a pathological increase in precision — the system becomes overly sensitive to noise, and phantom percepts emerge. The same framework explains both normal operation and the failure mode.
Toggle Auditory cortex pathway to see the full ascending hierarchy. Each level operates at a characteristic timescale, broadly consistent with the known EEG frequency bands. This visualisation is pedagogical — a schematic of how hierarchical inference might be imagined along the auditory pathway:
SG (Spiral Ganglion): cell bodies of the auditory nerve. First relay from IHC to brain. Fastest dynamics.
CN (Cochlear Nucleus): first brainstem relay. Onset detection, spectral processing, temporal pattern extraction.
IC (Inferior Colliculus): midbrain integration hub where nearly all ascending auditory pathways converge.
MGB (Medial Geniculate Body): the auditory thalamus. Descending predictions from cortex have their strongest modulatory effect here. Analogous to the lateral geniculate nucleus for vision.
A1 (Primary Auditory Cortex): tonotopic map on the superior temporal gyrus. Processes syllables, phonemes, auditory objects. Sends descending predictions back through the entire hierarchy.
Tone slider: generates a pure tone that directly stimulates the basilar membrane at the characteristic frequency location. Sweep it to see the activation point move along the spiral. Microphone: ambient sound adds live activity. Amplifier gain: controls evidence accumulation rate.
Parr, T., Pezzulo, G. & Friston, K. (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press.
Friston, K. (2010). The free-energy principle: a unified brain theory? Nat. Rev. Neurosci., 11(2), 127-138.
Liberman, M., Gao, J., He, D. et al. (2002). Prestin is required for electromotility of the outer hair cell and for the cochlear amplifier. Nature, 419, 300-304.
Dallos, P. (2008). Cochlear amplification, outer hair cells and prestin. Curr. Opin. Neurobiol., 18, 370-376.
Guinan, J. J. (2006). Olivocochlear efferents: anatomy, physiology, function, and the measurement of efferent effects in humans. Ear Hear, 27, 589-607.
Greenwood, D. D. (1990). A cochlear frequency-position function for several species. J. Acoust. Soc. Am., 87, 2592-2605.
Bastos, A. M. et al. (2015). Visual areas exert feedforward and feedback influences through distinct frequency channels. Neuron, 85(2), 390-401.
Computational method: CRR temporal grammar (temporalgrammar.ai) · Active Inference Institute · 2026