# Self-replicating artificial neural networks give rise to universal evolutionary dynamics

## TL;DR

A new deep-learning model called SeRANN can evolve like living organisms by copying itself and adapting over generations, showing how evolution can emerge in artificial systems.

## Contribution

SeRANN introduces endogenous mutations and implicit fitness in a self-replicating neural network model, enabling natural evolutionary dynamics.

## Key findings

- SeRANNs evolved over 6,000 generations showed adaptation, clonal interference, and epistasis.
- The model demonstrated evolution of mutation rates and distribution of fitness effects.
- Evolutionary phenomena similar to those in microbes were observed in the artificial system.

## Abstract

In evolutionary models, mutations are exogenously introduced by the modeler, rather than endogenously introduced by the replicator itself. We present a new deep-learning based computational model, the self-replicating artificial neural network (SeRANN). We train it to (i) copy its own genotype, like a biological organism, which introduces endogenous spontaneous mutations; and (ii) simultaneously perform a classification task that determines its fertility. Evolving 1,000 SeRANNs for 6,000 generations, we observed various evolutionary phenomena such as adaptation, clonal interference, epistasis, and evolution of both the mutation rate and the distribution of fitness effects of new mutations. Our results demonstrate that universal evolutionary phenomena can naturally emerge in a self-replicator model when both selection and mutation are implicit and endogenous. We therefore suggest that SeRANN can be applied to explore and test various evolutionary dynamics and hypotheses.

Computational self-replicators are often modeled with implicit definitions of fitness, using various approaches such as biophysical modelling, molecular dynamics, ecological dynamics, and computational logic. However, mutation is usually explicitly defined using standard probability distributions such as the Poisson distribution. Here, we attempted to develop a computational self-replicator with implicit fitness and mutation processes using artificial neural networks, which are popular, complex, and flexible models for approximating solutions to a wide range of computational problems. The result is the self-replicating artificial neural network (SeRANN), which learns to copy its own genotype while simultaneously solving a computational problem such as image classification. Approximation errors implicitly introduce mutations, which affect the source code of the networks, changing network hyper-parameters (e.g, number of neurons in a layer), network architecture (connections between layers), and hyper-parameters of the learning algorithm. We evolved a population of SeRANNs and observed various evolutionary phenomena often seen in evolutionary experiments with microbes, demonstrating that this new evolutionary framework provides a promising model for further studies in evolutionary theory.