alphadecaying

Alphadecaying is not a standard term in most scientific literature. It is sometimes used informally to refer to two related concepts that involve decay processes associated with the symbol alpha: alpha decay in nuclear physics and the decay of the learning rate parameter, often denoted alpha, in iterative optimization and machine learning. Because the word is ambiguous, documentation typically clarifies the intended meaning in context.

In nuclear physics, alpha decay describes the process by which a heavy atomic nucleus emits an alpha particle (two protons and two neutrons), resulting in a new nucleus with atomic number reduced by two and mass number reduced by four. The process occurs through quantum tunneling through the Coulomb barrier and is characterized by a half-life that varies widely across isotopes. Alpha decay is a key mechanism in the natural radioactivity of heavy elements such as uranium and radium and is exploited in dating methods and nuclear medicine, among other applications.

In optimization and machine learning, alpha commonly represents a learning rate or step size. An alphadecaying schedule reduces the learning rate as training progresses, often to improve convergence and reduce oscillations. Common forms include exponential decay, where alpha_t = alpha_0 exp(-kt); step or time-based decay, where alpha_t decreases according to a predefined schedule; and cosine annealing, which gradually lowers the learning rate toward a minimum. Decay can help stabilize training, but setting the schedule requires balancing exploration and convergence speed.

See also alpha decay, learning rate decay, and optimization scheduling.