Connectome-Constrained Artificial Neural Networks
Abstract
In biological neural networks (BNNs), structure provides a set of guard rails by which function is constrained to solve tasks effectively, handle multiple stimuli simultaneously, adapt to noise and input variations, and preserve energy expenditure. Such features are desirable for artificial neural networks (ANNs), which are, unlike their organic counterparts, practically unbounded, and in many cases, initialized with random weights or arbitrary structural elements. In this dissertation, we consider an inductive base case for imposing BNN constraints onto ANNs. We select explicit connectome topologies from the fruit fly (one of the smallest BNNs) and impose these onto a multilayer perceptron (MLP) and a reservoir computer (RC), in order to craft “fruit fly neural networks” (FFNNs). We study the impact on performance, variance, and prediction dynamics from using FFNNs compared to non-FFNN models on odour classification, chaotic time-series prediction, and multifunctionality tasks. From a series of four experimental studies, we observe that the fly olfactory brain is aligned towards recalling and making predictions from chaotic input data, with a capacity for executing two mutually exclusive tasks from distinct initial conditions, and with low sensitivity to hyperparameter fluctuations that can lead to chaotic behaviour. We also observe that the clustering coefficient of the fly network, and its particular non-zero weight positions, are important for reducing model variance. These findings suggest that BNNs have distinct advantages over arbitrarily-weighted ANNs; notably, from their structure alone. More work with connectomes drawn across species will be useful in finding shared topological features which can further enhance ANNs, and Machine Learning overall.