One-to-one mapping between deep network units and real neurons uncovers a visual population code for social behavior

Abstract

The rich variety of behaviors observed in animals arises through the complex interplay between sensory processing and motor control [1, 2, 3, 4, 5]. To understand these sensorimotor transformations, it is useful to build models that predict not only neural responses to sensory input [6, 7, 8, 9, 10] but also how each neuron causally contributes to behavior [11, 12]. Here we demonstrate a novel modeling approach to identify a one-to-one mapping between internal units in a deep neural network and real neurons by predicting the behavioral changes arising from systematic perturbations of more than a dozen neuron types. A key ingredient we introduce is “knockout training”, which involves perturbing the network during training to match the perturbations of the real neurons during behavioral experiments. We apply this approach to model the sensorimotor transformation of Drosophila melanogaster males during a complex, visually-guided social behavior [13, 14, 15, 16]. Contrary to prevailing views [17, 18, 19], our model suggests that visual projection neurons at the interface between the eye and brain form a distributed population code that collectively sculpts social behavior. Overall, our framework consolidates behavioral effects elicited from various neural perturbations into a single, unified model, providing a detailed map from stimulus to neuron to behavior.

Visual and motor signatures of locomotion dynamically shape a population code for feature detection in Drosophila

“…We note, however, that how downstream circuits make use of the information available across VPNs to guide behavior is not well understood, and further work incorporating connectomics, targeted perturbations of VPN channels, and behavioral analyses might shed light on this question. For example, a recent study used genetic silencing, coupled with a goal-oriented neural network model, to show that VPNs jointly encode behaviorally relevant visual features during Drosophila courtship ( Cowley et al, 2022 ).…”

Neuronal “parts list” and wiring diagram for a visual system

“…4b, S7). These VPNs are all looming detectors (Klapoetke et al 2022), and implicated in processing motion cues during social behavior (Cowley et al 2022). In short, our work indicates that LCs important for male courtship behavior receive information about color as well as motion and form.…”

Looming detection in complex dynamic visual scenes by interneuronal coordination of motion and feature pathways

“…On the other hand, Artificial Neural Networks (ANNs) trained on datasets can perform efficiently with numerous parameters but lack interpretability. Recently, neural scientists have started constraining ANNs with real neural circuit anatomy 30,34,35 , showing that such models can converge to biological solutions and be more efficient and robust ? .…”