ElegansNet: a brief scientific report and initial experiments

TL;DR

The paper investigates whether real neuronal connectomes can inform deep learning topology. It converts the Caenorhabditis elegans connectome into an almost 1:1 tensor-network architecture that maps sensors, interneurons, and motors to input, latent, and output spaces, respectively. The authors demonstrate that connectome-inspired networks outperform randomly wired counterparts on CIFAR-10 and show competitive unsupervised MNIST reconstruction results, highlighting the role of conserved topological features such as small-world structure. These findings suggest a principled, bio-inspired design path for robust and efficient deep learning systems that bridge neuroscience and AI.

Abstract

This research report introduces ElegansNet, a neural network that mimics real-world neuronal network circuitry, with the goal of better understanding the interplay between connectome topology and deep learning systems. The proposed approach utilizes the powerful representational capabilities of living beings' neuronal circuitry to design and generate improved deep learning systems with a topology similar to natural networks. The Caenorhabditis elegans connectome is used as a reference due to its completeness, reasonable size, and functional neuron classes annotations. It is demonstrated that the connectome of simple organisms exhibits specific functional relationships between neurons, and once transformed into learnable tensor networks and integrated into modern architectures, it offers bio-plausible structures that efficiently solve complex tasks. The performance of the models is demonstrated against randomly wired networks and compared to artificial networks ranked on global benchmarks. In the first case, ElegansNet outperforms randomly wired networks. Interestingly, ElegansNet models show slightly similar performance with only those based on the Watts-Strogatz small-world property. When compared to state-of-the-art artificial neural networks, such as transformers or attention-based autoencoders, ElegansNet outperforms well-known deep learning and traditional models in both supervised image classification tasks and unsupervised hand-written digits reconstruction, achieving top-1 accuracy of 99.99% on Cifar10 and 99.84% on MNIST Unsup on the validation sets.

Figures (2)

• Figure 1: The connectome of C.elegans is represented as a fully connected graph with two overlapping layers, where the solid edges represent chemical and directional synapses and the dashed edges represent electrical and undirected ones. The sensor neurons are represented in blue (Box (a)), while interneurons are represented in red (Box (b)). Finally, the motor neurons are represented in green (Box (c)). The encoder (Box (c)) may differ for various deep learning tasks and is responsible for generating the feature maps that serve as input to the sensor-tensors of the latent space. The learnable network (Box (d)) of tensors is the resulting layered model from the reference graph, projected into the center of the external environment. The tensor network takes the form of a directed acyclic graph and serves as the latent space of our models. In the output of the tensor network, the motor unit tensors are collected, and the most important features are selected in the decoder part (Box (d)).
• Figure 2: In this Figure, the original neural circuitry forming the ElegansNet$M1$ and $M2$ models is compared to randomly rewired models on two well-known problems: Cifar10 and MNIST Unsup, which are shown in boxes (a) and (b), respectively. The $y$-axis represents the validation set top-1 accuracy, while the $x$-axis represents the number of training epochs. The models derived from the tensor network of the original connectome are depicted by the red dashed lines, while the models built with randomly rewired small-world graphs (a total of 30) are represented by means of average and variance as follows: $\mathbf{G}_1$ (blue lines) is for Watts-Strogatz, $\mathbf{G}_2$ (orange dotted lines) is for Erdos-Renyi, and $\mathbf{G}_3$ (green dotted lines) is for Barabasi-Albert generators.