Flexible boolean computation by auditory neurons

Neurons have rich input-output functions for pooling and transforming inputs, owing to dendritic nonlinearities and adaptation. These nonlinearities are also flexible, adapting to accomodate changes in stimulus statistics (Młynarski and Hermundstad, 2021) or to extract different features from sensory space (Deny et al., 2017). Thus far, our understanding of the complex transformations by neurons has been largely informed by experiments which directly activate synaptic inputs on dendrites. However, these do not consider how natural and complex stimuli map to neuronal inputs. Our study uses ethologically-relevant and real-world sensory stimuli to probe neuronal integration functions. In auditory cortex in mice, we record extracellular single unit spikes to pairs of ultrasonic mouse vocalization (USV) syllables, which were selected and combined according to the neuron's responses. Many stimulus pairs produced a summation response (quantified by the summation index, SmI) that resembled the MAX operation: a pair of superimposed syllables evoked a response that was similar to the response to each syllable presented in isolation, instead of a linear summation, in agreement with Kozlov and Gentner (2014). To investigate the changes in integration behaviour to simple input transformations, we pitch-shifted syllables of a pair. Summation did not depend on the spectral separation between syllables in a pair. Instead, summation became more OR-like for pitch-shifted versions (of the same syllable pair) that were preferred by the neuron. When summation was more AND-like, we observed an increase in spike temporal precision to superimposed syllables, which could indicate increased selectivity. This flexibility in neuronal computation contrasts with neural activation functions artificial networks, which are limited to a prescribed set and fixed across units and time. Recent work has suggested that adopting flexible biological nonlinear activation functions improve task performance of artificial neural networks (Geadah et al, 2020). Examining how single neurons multiplex different computations may have interesting functional implications for network computation.