P064 Anatomically and Functionally Constrained Bio-Inspired Recurrent Neural Networks Outperform Traditional RNN Models
Mo Shakiba1,2, Rana Rokni1,2, Mohammad Mohammadi1,2,Nima Dehghani2,3*
1Neuromatch Academy, Neuromatch, Inc.
2N3HUB Initiative, Massachusetts Institute for Technology, Cambridge, U.S.A.
3McGovern Institute for Brain Research, Massachusetts Institute for Technology, Cambridge, U.S.A.
*Email: nima.dehghani@mit.edu
Introduction
Understanding how neural circuits drive sensory processing and decision-making is a central neuroscience challenge. Traditional Recurrent Neural Networks (RNNs) capture temporal dynamics but fail to represent the structured synaptic architecture seen in biological systems [1]. Recent spatially embedded RNNs (seRNNs) add spatial constraints for better biological relevance [2], yet they do not fully exploit detailed anatomical and functional data to enhance task performance and neural alignment.
Methods
We introduce a bio-inspired RNN that integrates detailed anatomical connectivity and two-photon calcium imaging data from the MICrONS dataset (https://www.microns-explorer.org/cortical-mm3), which offers nanometer-scale reconstructions and functional recordings from mouse visual cortex. Using neuronal positions, synaptic connections, functional correlations, and Spike Time Tiling Coefficients (STTC) [3]—a robust metric that eliminates firing rate biases—we constrain our model with biologically informed weight initialization, communicability calculations, and a regularizer that penalizes long-distance connections while boosting communicability to promote realistic network properties.
Results
Trained on three distinct decision-making tasks—a 1-step inference task, a Go/No-Go task, and a perceptual decision-making task— our bio-inspired RNN demonstrated significant performance improvements over baseline models across 30 simulations per model (900 total simulations across all model variants). Variants combined W* (biologically initialized weights) or W (standard initialization), D* (actual neuron distances) or D (random distances), and C (communicability calculation). Specifically, the anatomically and functionally constrained model (W*D*C) achieved the highest average accuracy across all tasks: 89.4% on the 1-step inference task, 96.9% on the Go/No-Go task, and 86.7% on the perceptual decision-making task.
Moreover, the biologically constrained model demonstrated superior performance across other evaluation metrics, including validation accuracy, training and validation loss, and network properties such as modularity and small-worldness. Specifically, the average modularity of the W*D*C and WD*C models was highest across all tasks, with values of 0.583 (1-Step Inference), 0.558 (Go/NoGo), and 0.594 (Perceptual Decision Making). Similarly, the average small-worldness was also the highest across two tasks, with values of 3.513 (Go/NoGo), and 4.325 (Perceptual Decision Making) (Fig. 1c-e)
Discussion
Our findings demonstrate that incorporating biological constraints into RNNs significantly boosts both task performance and the emergence of realistic network properties, mirroring actual neural architectures. Future work should extend this approach to visual processing tasks, explore other architectures such as LSTMs and GNNs, and integrate additional biological constraints.
Figure 1. (a) Weight initialization matrix (top left) from MICrONS data, combining functional correlation (bottom left) and STTC (bottom right) with log-normal noise. Anatomical distance matrix (top right) shows neuron positioning. (b) Top 10% models (900 simulations): Effects of λ and W on accuracy, loss, modularity, and small-worldness. (c-e) Model variants task performance shows WDC outperforming RNNs.
Acknowledgements
N.D. is supported by NIH Grant R24MH117295. The authors thank NIH for sponsoring DANDI archive, which provided the open-access data used in this study. M.S., R.R., and M.M. thank Neuromatch Academy for its support and resources for young scholars and this study. They also thank the DataJoint team for their help and guidance.
References
● Perich, M. G., & Rajan, K. (2020). Rethinking brain-wide interactions through multi-region ‘network of networks’ models. Current Opinion in Neurobiology, 65, 146–151.
● Achterberg, J., et al. (2023). Spatially embedded recurrent neural networks reveal widespread links between structural and functional neuroscience findings. Nature Machine Intelligence, 5(12), 1369–1381.
● Cutts, C. S., & Eglen, S. J. (2014). Detecting Pairwise Correlations in Spike Trains: An Objective Comparison of Methods and Application to the Study of Retinal Waves. The Journal of Neuroscience, 34(43), 14288–14303.
