Neuroevolution gives rise to more focused information transfer compared to backpropagation in recurrent neural networks

Abstract

Artificial neural networks (ANNs) are one of the most promising tools in the quest to develop general artificial intelligence. Their design was inspired by how neurons in natural brains connect and process, the only other substrate to harbor intelligence. Compared to biological brains that are sparsely connected and that form sparsely distributed representations, ANNs instead process information by connecting all nodes of one layer to all nodes of the next. In addition, modern ANNs are trained with backpropagation, while their natural counterparts have been optimized by natural evolution over eons. We study whether the training method influences how information propagates through the brain by measuring the transfer entropy, that is, the information that is transferred from one group of neurons to another. We find that while the distribution of connection weights in optimized networks is largely unaffected by the training method, neuroevolution leads to networks in which information transfer is significantly more focused on small groups of neurons (compared to those trained by backpropagation) while also being more robust to perturbations of the weights. We conclude that the specific attributes of a training method (local vs. global) can significantly affect how information is processed and relayed through the brain, even when the overall performance is similar.

Using artificial intelligence to fit, compare, evaluate, and discover computational models of decision behavior

“…Information theoretic approaches complement these correlational analysis, and have been used to characterize cognitive and computational systems alike (Tononi, 2004;Marstaller et al, 2013;Tehrani-Saleh & Adami, 2020;Hintze & Adami, 2023), and can also be combined with perturbation analysis (Bohm et al, 2022;Hintze & Adami, 2022).…”

“…This is another place where evolved systems differ from those typically used in AI: real brains tend to be more modular and sparsely connected compared to artificial neural networks (Happel & Murre, 1994). Using evolutionary algorithms that specify the presence or absence of connections rather than just their weight (Hintze et al, 2017;Hintze & Adami, 2022), creating and testing modular systems that perform specific functions (Bryson, 2005;Ellefsen et al, 2015), and drawing on the structure of the brain and biological systems (W. Cox & Dean, 2014) are likely to lead to more effective neural networks as well as artificial systems that more closely resemble the human brain.…”

Robustness of Sparsely Distributed Representations to Adversarial Attacks in Deep Neural Networks

“…Previous research has indicated that representations, or the information a neural network possesses about its environment, are dispersed throughout deep-learned networks [ 23 , 24 , 25 , 26 ]. In contrast, neural networks optimized using genetic algorithms do not exhibit this tendency and demonstrate greater robustness to noise [ 24 , 27 , 28 ]. Likewise, human brains, which have been “optimized” by evolution, employ distinct brain regions for computation [ 29 ].…”