P126 Predictive Coding in the Drosphila Optic Lobe
Rintaro Kai*1, Naoya Nishiura*1, Keisuke Toyoda*1, Masataka Watanabe*1 1The University of Tokyo Introduction
In recent years, the complete connectome of the fruit fly has been revealed [1], and its estimation of synaptic efficacy via backprogation training has lead to the reconstruction of T4/T5 motion selective cells [2]. However, in this particular study [2], biologically non-available optical flow was provided as a vector teaching signal. In this study, we used the complete connectome of the fruit fly and implemented Predictive Coding [3] by calculating the error between two tightly coupled cells,namely, L1 and C3. The results demonstrate the potential of training the full connectome neural circtuitry only using biological available vector teaching signals, say, the sensory input itself. Methods From the FlyWire dataset, we extracted connectivity information for neurons and 2,700,000 synapses in the right optic lobe and created a single-layer RNN and the neurotransmitters present at each synapse. The output function of each neuron was clipped, and the weights were normalized per post-neuron. Photoreceptor neurons received simulated natural video stimuli based on the shape of the fruit fly's eyes, then stimuli propagated to next neurons with each timestep. The network was trained using the mean squared error of outputs from anatomically close L1 and C3 neurons, creating a simple autoencoder using Predictive Coding [3]. Additionally, the activity of neurons at each timestep was visualized to ensure appropriate behavior. Results The learning of the task was successful, and the error converged to a very low value. Neurons other than those used for error calculation also showed appropriate activity, indicating that the network functioned effectively as a whole. Parameters were tuned effectively for the modeling settings, as phenomena where neuron outputs become constant regardless of input were also observed depending on the parameters.
Discussion The results of this study show that it is possible to perform unsupervised learning on the full connectome by taking errors between pairs of neurons, without incorporating artificial neurons or circuits.Future prospects include verifying whether the neuronal activity of the obtained model is biologically valid, examining the biological significance of hyperparameters, and testing whether network behavior and neuron role distribution can be robustly replicated compared to random initialization.
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] Dorkenwald, Sven et al. (2024). Neuronal wiring diagram of an adult brain. Nature, 634(8032), 124-138. [2] Lappalainen, Janne K. et al. (2024). Connectome-constrained networks predict neural activity across the fly visual system. Nature, 634(8036), 1132-1140. [3] Rao, Rajesh P. N. & Ballard, Dana H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79-87.
