Dice Question Streamline Icon: https://streamlinehq.com

Behavior at the critical “pentuple point” α = β = 1/2

Characterize the behavior of the loss curve and compute-optimal frontier at the critical point α = β = 1/2 in the (α, β)-phase diagram for the PLRF model trained with one-pass SGD, where all forcing and kernel components (Fpp, Fac, F0, Kpp) mix and interact.

Information Square Streamline Icon: https://streamlinehq.com

Background

The authors map out four phases (+ subphases) of compute-optimal behavior depending on data complexity α and target complexity β. Away from phase boundaries, distinct terms in the forcing and kernel functions dominate the loss and determine scaling exponents.

At the critical point α = β = 1/2—dubbed the pentuple point—all components contribute and interact, making the dynamics nontrivial. The authors highlight this as an unresolved case where a complete characterization remains to be established.

References

Moreover, there exists an interesting critical point \alpha = \beta = \tfrac{1}{2} where all the parts of the forcing function and kernel mix and interact with each other. The behavior of the loss at the pentuple point (see Fig~\ref{fig:phase_diagram}) we leave for future research.

4+3 Phases of Compute-Optimal Neural Scaling Laws (2405.15074 - Paquette et al., 23 May 2024) in Section “The 4 Phases”