SGLD favors wide loss minima through the Eyring free-energy formula, and GANs act like a predator-prey system that pushes learning out of narrow likelihood maxima.
Towards physical principles of biological evolution
1 Pith paper cite this work. Polarity classification is still indexing.
abstract
Biological systems reach organizational complexity that far exceeds the complexity of any known inanimate objects. Biological entities undoubtedly obey the laws of quantum physics and statistical mechanics. However, is modern physics sufficient to adequately describe, model and explain the evolution of biological complexity? Detailed parallels have been drawn between statistical thermodynamics and the population-genetic theory of biological evolution. Based on these parallels, we outline new perspectives on biological innovation and major transitions in evolution, and introduce a biological equivalent of thermodynamic potential that reflects the innovation propensity of an evolving population. Deep analogies have been suggested to also exist between the properties of biological entities and processes, and those of frustrated states in physics, such as glasses. We extend such analogies by examining frustration-type phenomena, such as conflicts between different levels of selection, in biological evolution. We further address evolution in multidimensional fitness landscapes from the point of view of percolation theory and suggest that percolation at level above the critical threshold dictates the tree-like evolution of complex organisms. Taken together, these multiple connections between fundamental processes in physics and biology imply that construction of a meaningful physical theory of biological evolution might not be a futile effort.
citation-role summary
citation-polarity summary
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Control of Overfitting with Physics
SGLD favors wide loss minima through the Eyring free-energy formula, and GANs act like a predator-prey system that pushes learning out of narrow likelihood maxima.