ABSTRACT
This paper investigates breakpoints in scaling laws for complex system identification, proposing a methodology to detect and interpret these transitions, drawing parallels to phase transitions in statistical physics.
PAPER · PDF
Loading PDF…
Key findings
Breakpoints in scaling laws indicate sudden transitions in learning behavior.
Proposed framework integrates changepoint detection with scaling law analysis.
Experiments span multiple benchmark systems, architectures, and scaling dimensions.
Limitations & open questions
Breakpoint detection may be affected by measurement noise.
The generalizability of findings to other domains is yet to be determined.