UBA
An open-loop learning rate decay design framework based on the total training budget, which drives the schedule by computing the optimal learning rate curve offline anchored to iteration count; however, it relies on SGD dynamics and its optimization difficulty and parameter meanings are ambiguous.
Timeline (last 90 days)
UBA (Universal Budget Aware) learning rate decay relies on a fixed offline trajectory, anchored only to the iteration count, and cannot adjust dynamically based on the model's internal state, so it is essentially open-loop control.
The derivation of UBA is built on the continuous-time dynamics of SGD, so its optimality conclusion holds for SGD but is difficult to transfer directly to adaptive optimizers such as Adam.
All Facts (2)
UBA (Universal Budget Aware) learning rate decay relies on a fixed offline trajectory, anchored only to the iteration count, and cannot adjust dynamically based on the model's internal state, so it is essentially open-loop control.
50%UnverifiedThe derivation of UBA is built on the continuous-time dynamics of SGD, so its optimality conclusion holds for SGD but is difficult to transfer directly to adaptive optimizers such as Adam.
50%