Meeting room: https://tu-dresden.zoom.us/j/97802507999?pwd=QTFUNnhXRmZwbnZhN2U3R1lhWG9EUT09 (room 97802507999, psw: cns-2020)
Alejandro Tabas,
Glad Mihai,
Stefan Kiebel,
Robert Trampel,
Katharina von KriegsteinSensory pathways efficiently transmit information by adapting the neural responses to the local statistics of the sensory input. The predictive coding framework suggests that sensory neurons constantly match the incoming stimuli against an internal prediction derived from a generative model of the sensory input. Although predictive coding is generally accepted to underlay cortical sensory processing, the role of predictability in subcortical sensory coding is still unclear. Several studies have shown that single neurons and neuronal ensembles of the subcortical sensory pathway nuclei exhibit stimulus specific adaptation (SSA), a phenomenon where neurons adapt to frequently occurring stimuli (standards) yet show restored responses to a stimulus with deviating characteristics from the standard (deviant). Although neurons showing SSA are often interpreted as encoding prediction error, computational models to date have successfully explained SSA in terms of local network effects based on synaptic fatigue. Here, we first introduce a novel experimental paradigm where abstract rules are used to manipulate predictability. 19 human participants listened to sequences of pure tones consisting on seven standards and one deviant while we recorded mesoscopic responses in auditory thalamus and auditory midbrain using 7-Tesla functional MRI. In each sequence, the deviant was constrained to occur once and only once, and always in locations 4, 5 or 6. Although the three locations were equiprobable at the beginning of the trial, the conditional probability of hearing a deviant in location n after hearing n-1 standards is 1/3, 1/2, and 1, for deviant locations 4, 5, and 6, respectively. This paradigm yields different outcomes for habituation and predictive coding: if adaptation is driven by local habituation only, the three deviants should elicit similar neuronal responses; however, if it is predictive coding that entails adaptation, the neuronal responses to each deviant should depend on their abstract predictability. Our data showed that the responses to the deviants were strongly driven by abstract expectations, indicating that predictive coding is the main mechanism underlying mesoscopic SSA in the subcortical pathway. These results are robust even at the single-subject level. Next, we developed a new model of pitch encoding for pure tones following the main directives of predictive coding. The model comprises two layers whose dynamics reflect two different levels of abstraction. The lower layer receives its inputs from the auditory nerve and makes use of the finite bandwidth of the peripheral filters to decode pitch fast and robustly. The second layer holds a sparse representation that integrates the activity in the first layer only once the pitch decision has been made. Top-down afferents from the upper layer reinforce the pitch decision and accept the inclusion of priors that facilitate the decoding of predictable tones. Without the inclusion of priors, the model explains the key elements of SSA in animal recordings at the single-neuron level, as well as the main phenomenology of its mesoscopic representation. The inclusion of priors reflecting the abstract rules described in our paradigm facilitates the decoding of tones according to their predictability, effectively modulating the responses at the mesoscopic level. This modulation affects the mesoscopic fields generated during pitch encoding, fully explaining our experimental data.
