How Does the Fruit Fly’s Brain See the World?

Scientists have built a new model to decode how fruit‑fly neurons react to what they see. It’s called Multi‑Path Aggregation (MPA) and relies on steady‑state neural network theory. The trick is to merge visual inputs with the fly’s complete brain wiring diagram, the connectome. MPA computes powers of the adjacency matrix and optimizes finite paths, quickly predicting each neuron’s response. The output is a Visual Function Profile (VFP) for every cell, indicating ON/OFF activation, direction preference, and reactions to complex stimuli. The model achieved a Pearson correlation of 0.84 for ON/OFF responses, far surpassing earlier approaches. Moreover, researchers replaced conventional CNN modules with VFP‑derived lobula columnar responses, enabling a simulated fly to navigate and dodge obstacles. This shows the model not only explains internal processing but also drives real‑world behavior. In short, MPA provides a functional map linking connectome, neuron properties, and behavior, paving the way for brain‑inspired intelligence.

The Multi‑Path Aggregation framework rests on a linear formulation of activity flow through the synaptic network. Starting from the connectome adjacency matrix A, the model computes A^k for k = 1…K, where K is the maximum number of steps considered. Each power captures pathways of length k, allowing the model to incorporate both feed‑forward and feedback effects across multiple layers. Optimization includes an L2 regularization term to prevent over‑fitting and an L1 sparsity term that favors biologically plausible routes. Visual input is encoded as a spatial feature vector (luminance, contrast, orientation) extracted by Gabor‑like filters mimicking the fly’s retina. This vector is projected into neuronal space via a weight matrix W, obtained through ridge regression on electrophysiological response data. The final response of neuron i is r_i = σ( Σ_{k=1}^K α_k (A^k W x)_i ), where σ is a sigmoid activation and α_k are learned coefficients weighting each path order’s importance. Training employs gradient descent with back‑propagation on a dataset of ON, OFF, and directional stimulus recordings. Cross‑validation demonstrates robust generalization to novel stimuli, reducing mean‑squared error by 45% compared to conventional convolutional networks. Moreover, selective synaptic block simulations reveal that removing specific higher‑order pathways significantly impairs navigation, highlighting the crucial role of long‑range connections in the fly’s visual computation. This approach paves the way for AI models that exploit connectome topology to optimize sensory functions, cutting down on parameter count and training data requirements.