Po-Ting Lin (林柏廷)

Work in motion

Research Log

A running ledger of my research projects — the ones still moving, and the closing notes on the ones that finished. Active projects may be just a title and a few keywords; concluded ones come with a reflection on how they landed.

In progress

  • Cross-View Variance Field

    A stray observation in path-traced stereo noise, chased until it became a measurable matching shortcut and a preprint — with the harder question still open.

    • path tracing
    • stereo matching
    • variance correlation
    • synthetic data
    In progress

  • Structural Crystallization

    The unifying theory behind most of my other work — a computational framework separating memory's structural accumulation from its representational fidelity. Declined at Psychological Review with substantive expert reviews; now under revision, and actively being extended into running systems.

    • structural crystallization
    • memory
    • computational theory
    • emergence
    In progress

  • Crystalmem

    In progress

  • Intrinsic Pixel Complexity

    In progress

  • Per-Object Variance Fingerprint

    In progress

  • PRISM

    Paused

  • StaMask

    Follow-up to PIDS: BatchNorm and L1 normalization in standard stereo architectures appear to erase the transparent-surface variance signal before it reaches the matching layer. Wider implications still open — no final verdict yet.

    • stereo architectures
    • normalization
    • BatchNorm
    • variance signal
    Paused
Concluded

  • PIDS — Physics-Informed Deep Stereo

    A five-month, 45-experiment investigation into polarized stereo for transparent-object depth — closed on proving the path is an optical dead end, which is itself the result.

    Concluded Feb 2026

    Concluded

  • CrystalASR

    A modular, phoneme-grounded speech recognizer built to test whether upward information flow through explicit linguistic structure beats end-to-end opacity.

    Concluded

  • CrystalCache

    Long-context KV-cache eviction designed by porting a two-dimensional model of biological memory into a scoring formula.

    Concluded

  • PIRNet

    Concluded

  • TF32 Cross-Architecture Study

    A negative-result investigation into a 33% accuracy gap between H200 and RTX 4090 — and the falsification of my own first hypothesis.

    Concluded