Seeking an “Effective Loss” for Grokking in Modular Addition
A checkpoint-grounded case study on modular-addition grokking: Fourier features, phase alignment, global-phase failure modes, and confound-aware score analysis.
hi i'm chryseis liu. depending on how you found this site, you might know me as a researcher or a mediocre musician or an aspiring mediocre athlete. i'm now interested in the 'Physics of AI', macroscopic & beyond-reductionism interpretability, complex systems, and emergence.
what's been living rent-free in my head on walks and in showers lately : epiplexity, compositionality of language (as opposed to image), statistical physics, inductive biases implied by architectures, philosophy of science & epistemology, and how the heck we're in a historically unusual place where we can measure almost anything and explain almost nothing.
A checkpoint-grounded case study on modular-addition grokking: Fourier features, phase alignment, global-phase failure modes, and confound-aware score analysis.
2025 is a mathematically rare year as it's a perfect square (45²). The last perfect square year was 1936 (44²) and the next will be 2116 (46²), so this is most likely the only perfect square year that any of us will ever experience in our prime.
My shore is low, so frail and thin
Six inches more, and death will flood in
unpolished/unedited dump
click cover photos to view collections • click individual photos for full-size view