The Stacked Autoencoder Evolution Hypothesis

TL;DR

The paper addresses how evolution can exhibit both gradual and punctuated changes that are not fully captured by traditional mutation–selection models. It proposes the Stacked Autoencoder Evolution Hypothesis, positing that self-replication internally compresses and reconstructs genetic information across multiple hierarchical latent spaces, enabling mutations at deeper layers to drive coordinated, large-scale phenotypic shifts. The authors illustrate the plausibility of this mechanism with an artificial-chemistry framework and a minimal neural-network–based simulation, showing emergent hierarchical encoding/decoding structures and cross-layer information preservation. While the model is intentionally abstract and not a direct biological claim, it offers testable predictions about hierarchical information processing in evolution and suggests a bridge to representation-learning concepts in computational biology. The framework provides a novel lens to understand both continuous and discontinuous evolutionary transitions and could inform synthetic molecular design and future empirical exploration.

Abstract

This study introduces a novel theoretical framework, the Stacked Autoencoder Evolution Hypothesis, which proposes that biological evolutionary systems operate through multi-layered self-encoding and decoding processes, analogous to stacked autoencoders in deep learning. Rather than viewing evolution solely as gradual changes driven by mutation and selection, this hypothesis suggests that self-replication inherently compresses and reconstructs genetic information across hierarchical layers of abstraction. This layered structure enables evolutionary systems to explore diverse possibilities not only at the sequence level but also across progressively more abstract layers of representation, making it possible for even simple mutations to navigate these higher-order spaces.Such a mechanism may explain punctuated evolutionary patterns and changes that can appear as if they are goal-directed in natural evolution, by allowing mutations at deeper latent layers to trigger sudden, large-scale phenotypic shifts. To illustrate the plausibility of this mechanism, artificial chemistry simulations were conducted, demonstrating the spontaneous emergence of hierarchical autoencoder structures. This framework offers a new perspective on the informational dynamics underlying both continuous and discontinuous evolutionary change.

Figures (7)

• Figure 1: Schematic representation of a stacked autoencoder neural network. Red connections represent the encoding weights used to compress the input information, while blue connections represent the decoding weights used to reconstruct the input from the internal representations. Training is performed layer-wise.
• Figure 2: Self-replication conceptualized as a chain of molecular transformations. Intermediate molecules implement compression and reconstruction of genetic information, producing abstract latent representations analogous to autoencoder architectures.
• Figure 3: Mutations introduced at different layers of the replication hierarchy influence either physical sequences (shallow layers) or abstract latent representations (deep layers), with deeper perturbations being decoded into concrete molecular changes capable of producing discontinuous genetic novelty.
• Figure 4: Artificial chemical reaction network in simulation. $M_{\#}$ represents molecules of length #. Arrows indicate transformations from $M_{\#1}$ to $M_{\#2}$, where $M_{\#3}$ placed along the arrow serves as a catalyst. For example, the reaction in the upper left shows that a molecule $M_{13}^{i}$ is converted to a molecule $M_{7}^{j}$ through catalysis by another molecule $M_{13}^{l}$. Catalytic functions utilize convolutional networks (encoding) or deconvolutional networks (decoding). $f^{E}_{M_{\#1}}(M_{\#2})$ denotes a convolutional network constituted by $M_{\#1}$ that receives $M_{\#2}$ as input. The parameters $k$, $p$, and $s$ of the catalyst molecule represent kernel size, padding, and stride of convolutional or deconvolutional networks, respectively.
• Figure 5: Distributions of $M_{13}$, $M_{7}$, and $M_{3}$ molecular states at each step. The distributions are visualized in three dimensions by applying principal component analysis (PCA) for dimensionality reduction.