Research work · 2026
Algorithmic Mass: A Theory of Stability and Change in Machine Learning
Jincheng Zhang
Zenodo (CERN European Organization for Nuclear Research)
Abstract
This paper introduces the concept of "Algorithmic Mass," a theoretical framework for understanding the inherent stability and resistance to change within machine learning models. Algorithmic Mass is defined as a measure of the accumulated complexity and learned parameters within a model, directly proportional to its difficulty to modify. The theory posits that higher Algorithmic Mass leads to increased resistance to alterations, particularly through techniques like transfer learning, fine-tuning, and catastrophic forgetting. A core tenet of the theory is the pursuit of a unified explanation for observed model behavior, linking Algorithmic Mass to the fundamental limits of adaptability. We propose a formal definition of Algorithmic Mass and outline key relationships between this metric and various machine learning phenomena. This work lays the foundation for a new perspective on model design and optimization, prioritizing stability and predictable change management. ---