A model of functional recovery after significant loss of neural tissue: biofeedback based healing of vestibular dysfunction

Our model contains 4 populations of rate-coded units with sigmoid activation functions that are either not or fully connected via activity modulated Hebbian synapses (see Figure 1). A vestibular apparatus (VA) senses the tilt level of the modeled organism. VA is connected to a hidden population (HL) connected to a motor control population (BA), generating balancing actions and thereby closing a control loop by influencing the current tilt level. A second loop, the biofeedback, contains a population mimicking the signal of the mentioned tongue strip (TS).

VA and TS create population-coded output because their units are broadly tuned to different preferred tilt levels. HL and BA use winner take all dynamics. All units receive, in addition to the afferent input, a constant amount of white noise. Feedback connections from BA to HL force these populations to commit to a common, converged state.

Destroyed VA-units reduce the total input to HL. Homeostatic input normalization iteratively strengthens remaining postsynaptic processes to regain the desired input strength.

After destruction of a significant amount of VA-nodes (>90%) the remaining efferent signal does not exceed HL’s noise level and the entire system turns non-functional. During homeostatic input normalization the tuning of remaining efferent VA connections broadens and causes the system to settle in a non-functional state.

Biofeedback substitutes missing vestibular data and re-enables BA to generate sensible actions. BA-HL-feedback forces HL’s output to be correlated with the sensed tilt angles. Thus, activity modulated Hebbian learning re-sharpens VA’s efferent tuning and the modeled organism relearns to balance in an upright position – even without biofeedback. Phenomenologically this effect is also observed in human patients.