An Artificial Neural Network for Image Classification Inspired by Aversive Olfactory Learning Circuits in Caenorhabditis Elegans

TL;DR

The results indicate that the ANN inspired by the aversive olfactory learning circuits of C. elegans achieves higher accuracy, better consistency and faster convergence rates in image classification task, especially when tackling more complex classification challenges.

Abstract

This study introduces an artificial neural network (ANN) for image classification task, inspired by the aversive olfactory learning circuits of the nematode Caenorhabditis elegans (C. elegans). Despite the remarkable performance of ANNs in a variety of tasks, they face challenges such as excessive parameterization, high training costs and limited generalization capabilities. C. elegans, with its simple nervous system comprising only 302 neurons, serves as a paradigm in neurobiological research and is capable of complex behaviors including learning. This research identifies key neural circuits associated with aversive olfactory learning in C. elegans through behavioral experiments and high-throughput gene sequencing, translating them into an image classification ANN architecture. Additionally, two other image classification ANNs with distinct architectures were constructed for comparative performance analysis to highlight the advantages of bio-inspired design. The results indicate that the ANN inspired by the aversive olfactory learning circuits of C. elegans achieves higher accuracy, better consistency and faster convergence rates in image classification task, especially when tackling more complex classification challenges. This study not only showcases the potential of bio-inspired design in enhancing ANN capabilities but also provides a novel perspective and methodology for future ANN design.

Figures (4)

• Figure 1: a, Implementation principle of behavioral experiment on learning to avoid aversive olfactory stimuli in C. elegans. b, Fundamentals of high-throughput gene sequencing technology. c, Gene expression fold changes between learning and control cohorts. Positive values denote upregulated and negative values indicate downregulated. To minimize the measurement error, we limit the fold changes to fall within the range of [-50, 50]. d, Gene expression percentage in specific functional neurons, excluding functional neurons with expression percentage below 2%. e, Identification of functionally correlated neurons in aversive olfactory learning. f, The biological neural network of C. elegans. The weight of connection is proportional to the number of synapses $EW_{ij}$ between two functional neurons, while the intensity of color within the functional neurons indicates the magnitude of $CRI_{i}$.
• Figure 2: a, The steps of extending the functional neural circuits with sensory neuron as starting neuron. The bracketed value next to arrow line represents the rank determined by the number of synapses $EW_{ij}$ contained in this synaptic connection from large to small. b, The steps of extending the functional neural circuits with interneuron as intermediate neuron. c, The steps of extending the functional neural circuits with motor neuron as terminal neuron. d, The functional neural circuits associated with aversive olfactory learning of C. elegans, which consist of 22 functional neurons (10 sensory neurons, 5 interneurons and 7 motor neurons) and 21 synaptic connections.
• Figure 3: a, The framework map of an artificial neural network for image classification inspired by the neural circuits responsible for aversive olfactory learning in C. elegans. The part framed by green dotted line is the functional neural circuits module of nematode. b, The framework map of an artificial neural network for image classification inspired by the randomized neural circuits. The part framed by green dotted line is the randomized functional neural circuits module of nematode. c, The framework map of an artificial neural network for image classification modeled on LeNet lecunNet1998.
• Figure 4: The accuracy of classification across all categories on MNIST (a), FashionMNIST (b), CIFAR10 (c), CIFAR100 (d). The consistency of classification accuracy among different categories on MNIST (e), FashionMNIST (f), CIFAR10 (g), CIFAR100 (h). The rate of convergence of classification loss across all categories on MNIST (i), FashionMNIST (j), CIFAR10 (k), CIFAR100 (l).