KIP-Veröffentlichungen

Animals collect information about their environment through a variety of senses that need to be integrated into a coherent perspective. Since all sensory information is both incomplete and corrupted by noise, this integration has the main goal of increasing the information otherwise obtained from only a single sense. The Bayes-optimal estimate for the most likely stimulus under the assumption of independent Gaussian noise is obtained by averaging estimates from different modalities while weighting each with its respective reliability.

It was previously demonstrated in behavioral experiments that animals and humans combine multisensory stimuli in this optimal manner (Ernst and Banks, 2002; Fetsch et al., 2009; Nikbakht et al., 2018). What type of neuronal circuitry is able to perform such sensory integration? We present a neuron model capable of implementing the required computations by exploiting the biophysical dynamics of conductance-based neurons with dendritic compartments. Furthermore, a plausible error-driven plasticity rule enables neurons to learn not only input-output mappings, but to also simultaneously represent the respective reliabilities of each input that are necessary for a Bayes-optimal integration. In addition, the model supports dynamic reweighting of modalities and can thereby react to changes in stimulus reliabilities on a much shorter time scale than the one of synaptic plasticity. While both neuron and synapse dynamics are derived from a probabilistic description of neuronal processing, the model does not require a Bayes-optimal teacher but only input-output samples, allowing efficient learning.

To illustrate our model, we present a feed-forward circuit receiving input from two different modalities with different associated reliabilities and show that after learning, the circuit optimally takes into account the respective reliabilities when processing new information. Finally, we discuss extensions of our model to non-linear dendritic compartments and to multi-layered cortical circuits that learn continuous input-output mappings (Dold et al., 2018).