Connectome-based modeling provides a powerful framework to investigate the inner workings of the biological brain, but its supervised training often relies on external labels unavailable in real environments [1]. In the Drosophila visual system, prior work employed a connectome-constrained network with optical flow as teaching signals [2], whereas, neural circuits in Drosophila does not have access to such labels. To address this limitation, we adopted an autoencoder-like strategy: a network reconstructing R1–8 neural responses. We confirmed development of direction selectivity T4 and T5 cells, leveraging only the brain’s innate connectivity [3].
Methods We adopted the non-trained circuitry of the deep mechanistic network (DMN) of the Drosophila optic lobe [2], while removing the artificial network receiving optical flow as the teaching signal. Instead, we introduced a set of “phantom” R1–8 neurons that only receives feedback from the optical lobe. During training, the model’s sensory input was compared to the outputs of these phantom neurons via an L2 reconstruction loss. We preserved the native connectome structure, including the hexagonal columnar organization. Standard gradient-based optimization was used to update neuronal and synaptic parameters. Results After training, the DMN produced retinal-like activity patterns in its intermediate layers, effectively mapping spatial shadows across the hexagonal retinotopic array [4]. Notably, T4 neurons acquired direction-selective responses comparable to those observed in supervised settings, though preferred directions were not identical to biological measurements [2]. These results demonstrate that training of connectome-based autoencoder architecture leads to motion-selective T4 and T5 neurons, reproducing the functioning drosophila optical lobe. Discussion Our findings show that biologically plausible, connectome-constrained networks can self-organize fundamental visual computations through an autoencoder framework rather than providing explicit teaching signals [2]. By exploiting Drosophila’s neural connectivity and reconstructing phantom R1–8 neurons, the model reveals how intrinsic circuit architecture may lead to acquirement of direction selective cells. Our results illustrate the potential of training connectome networks under a biological plausible architecture, namely, auto-encoders, which may lead to near-ground-truth neural dynamics.
Acknowledgements This work has been supported by the Mohammed bin Salman Center for Future Science and Technology for Saudi-Japan Vision 2030 at The University of Tokyo (MbSC2030) and JSPS KAKENHI Grant Number 23K25257. References [1]https://doi.org/10.3389/fncom.2016.00094 [2]https://doi.org/10.1038/s41586-024-07939-3 [3]https://doi.org/10.7554/eLife.40025 [4]https://doi.org/10.1073/pnas.1509820112
