The Origin of Inference: Ediacaran Ecology and the Evolution of Bayesian Brains

TL;DR

This paper proposes a model explaining the evolution of nervous systems and spiking neurons in Ediacaran animals as adaptive decision-making tools for predator avoidance, demonstrating near-optimal Bayesian inference capabilities.

Contribution

It introduces a simple neural network model that performs Bayesian inference for escape decisions, linking neural architecture to evolutionary pressures in early animals.

Findings

Neural threshold devices act as utility-maximizing decision-makers.

A neural network can approximate Bayesian optimal decisions.

A subnetwork computes Bayesian posterior density for predator-prey distance.

Abstract

The evolution of spiking neurons and nervous systems in the late Ediacaran period simultaneously with the evolution of carnivores around 550 million years ago can be explained by the need for accurately timed decisions under an imminent threat of being eaten. A simple model shows that threshold triggering devices, spiking neurons, are utility-maximizing decision-makers for the timing of escape reflexes given the sensory cues available to Ediacaran animals at the onset of carnivory. Decisions are suboptimal for very weak stimuli, providing selection pressure for secondary processing of primary spike train data. A simple network can make approximately Bayes optimal decisions given stochastic spike trains. Decisions that are arbitrarily close to Bayes optimal can be obtained by enlarging this network. A subnetwork that computes the Bayesian posterior density of the critical state variable - distance between predator and prey - emerges as a core component of the decision-making mechanism. This is a neural analog of a Bayesian particle filter with cerebellar-like architecture. The model explains fundamental properties of neurons and nervous systems in modern animals and makes testable predictions.