7/ In the era of limited funding, our work showcases how to use models to bridge neural and natural behavior data to (1) increase discrimination power over neural models, (2) improve behavioral predictions, and (3) reveal novel bioplausible algorithms for neural computation in natural settings.

6/ Finally, we studied the nonlinear accumulation model of song encoding more closely, revealing previously unknown song patterns driving female slowing, and a neural algorithm for encoding long input sequences that leverages nonlinear adaptation to remember fine temporal patterns for long periods.

5/ Methodologically, our work shows how natural behavior can refine predictions of how neural data generalize beyond their original experimental context AND that modeling hidden neural activity can improve pure natural behavior predictions, even relative to popular black-box deep networks.

4/ This suggests that linear-nonlinear feature detection is not enough, but rather that flies may encode long communication sequences via nonlinear accumulation along multiple dimensions of activity space in a heterogeneous neural population code for song history.

3/ We found that one encoding model, based on multi-dimensional, nonlinear accumulation, allowed us to predict female locomotion much better than a classic linear-nonlinear feature-detection model, also outperforming many other predictors, including several black-box artificial neural networks.

2/ To gain further model discrimination power, we turned to a separate pure-behavior dataset of naturalistic fly courtship. We simulated the female’s neural responses to the male’s song using the encoding models then tried to predict her locomotion from the simulated neural data.

1/ How is the male fruit fly’s complex courtship song encoded in the female fly brain? Calcium imaging of responses to simplified song stimuli suggest neural codes are spread across a population with heterogeneous selectivities and timescales, but multiple encoding models fit the data equally well.