Towards a Theory of Evolution as Multilevel Learning (2110.14602v1)

Abstract: We apply the theory of learning to physically renormalizable systems in an attempt to develop a theory of biological evolution, including the origin of life, as multilevel learning. We formulate seven fundamental principles of evolution that appear to be necessary and sufficient to render a universe observable and show that they entail the major features of biological evolution, including replication and natural selection. These principles also follow naturally from the theory of learning. We formulate the theory of evolution using the mathematical framework of neural networks, which provides for detailed analysis of evolutionary phenomena. To demonstrate the potential of the proposed theoretical framework, we derive a generalized version of the Central Dogma of molecular biology by analyzing the flow of information during learning (back-propagation) and predicting (forward-propagation) the environment by evolving organisms. The more complex evolutionary phenomena, such as major transitions in evolution, in particular, the origin of life, have to be analyzed in the thermodynamic limit, which is described in detail in the accompanying paper.

• The paper proposes a unifying framework viewing evolution as multilevel learning that integrates renormalizability and learning dynamics.
• It employs mathematical models, including Fourier and wavelet transforms, to draw analogies between evolutionary fitness and neural network loss functions.
• The theory challenges traditional views by emphasizing multilevel selection as essential for complex phenomena like natural selection and the origin of life.

Overview of Evolution as Multilevel Learning

The paper "Towards a Theory of Evolution as Multilevel Learning" proposes a comprehensive theoretical framework that conceptualizes biological evolution through the lens of machine learning, specifically multilevel learning processes. This framework is presented with the goal of integrating the myriad of evolutionary phenomena within a single mathematical paradigm, highlighting the confluence of renormalizability, learning dynamics, and evolutionary principles.

Fundamental Principles and Correspondence with Learning

The authors delineate seven fundamental principles essential for an observable universe where evolution is feasible. Among these principles are the existence of a loss function, hierarchy of scales, frequency gaps, renormalizability, extension capabilities, replication, and asymmetric information flow. These principles align closely with learning theory as applied to neural networks, suggesting that biological systems can be analyzed using learning dynamics. The correspondence is striking; for instance, the loss function central to learning theory is proposed as an analogue to the fitness function in evolution, and multilevel selection is seen as endogenous to learning.

Theoretical Implications and Major Claims

The paper extends the analogy, proposing that biological evolution is inherently multilevel, with multilevel selection being an indispensable feature rather than a contentious hypothesis. This perspective challenges conventional views, highlighting the necessity of multilevel interactions to facilitate major evolutionary transitions, including the origin of life itself. The mathematical framework employing neural networks suggests that complex evolutionary phenomena such as natural selection, parasitism, and programmed death can be modeled accurately.

A bold claim evident in this work is the concept that the universe can wholly be described as a neural network, driving a broader "learning" process across all levels of evolution—from atomic particles to galaxy clusters. This assertion places the concept of learning as the underpinning mechanism of evolutionary complexity and diversifies the understanding of Darwinian natural selection.

Numerical Results and Observations

The authors provide mathematical models supporting these claims, primarily focusing on the loss function and its configuration in the learning processes. The use of Fourier and wavelet transforms exemplifies the potential to encapsulate evolutionary dynamics in mathematical terms, facilitating predictions within this learning framework. The theory predicts separation of scales and the rise of multilevel hierarchy naturally within learning systems.

Future Directions and Considerations

Speculatively, this framework allows us to rethink the origin of life as a result of complex system learning dynamics—a major evolutionary trend that is not necessarily unique but rare under specific environmental constraints. The paper offers avenues for further inquiry, particularly into the thermodynamic aspects of evolutionary processes, which are analyzed in an accompanying paper. Additionally, it suggests that adaptiveness and diversity, encapsulated in multilevel learning systems, might predict the emergence of complexity beyond our current understanding.

Conclusion

The theory proposes a unified perspective on biological evolution, embedding it within the general principles of learning, potentially altering the narrative from natural selection as an isolated process to a component of a multilevel learning system. This novel framework presents substantial conceptual utility, aiming to redefine traditional evolutionary biology while advocating for greater integration with physics and machine learning. While the paper outlines the theory succinctly, it prompts a broader examination and empirical validation within the realms of evolutionary studies and computational biology.