Mini Paper

Problem

在二维合成线性可分数据上,加入少量标签噪声后,带固定学习率的 perceptron 是否仍然能明显优于常数分类 baseline?

Core Idea

即使存在少量标签翻转,简单 perceptron 仍能在二维合成数据上取得明显高于 majority-class baseline 的 balanced accuracy;但其稳定性会随噪声上升而下降。

Method

  • Define the hidden lag variable k and a partially observed sequence x_t with local zeroed-out spans.
  • Construct a global non-parametric baseline by searching the lag that minimizes mean absolute disagreement between x_t and x_(t-k) on observed pairs.
  • Construct the local-to-global method by slicing the sequence into overlapping windows, estimating a local best lag in each window, and then aggregating these lag votes into one global prediction.
  • Treat degradation as part of the protocol: if the full method fails, drop to a smaller problem, then to baseline-only, then to a dry-run or failure note instead of leaving the day blank.

Experimental Setup

  • topic_id: manual-在二维合成线性可分数据上-加入少量标签噪声后-带固定学习率的-perceptro
  • strategy: full
  • case_count: 0
  • anomaly: False

Result

  • status: ok
  • accuracy: n/a

Interpretation

The unit stays valuable when the improved method fails because the baseline, the degradation path, and the final reasoning remain explicit.

Artifact References

  • task_id: research-2026-03-11b
  • result_summary_ref: agn://artifact/c0a8c848c9364a7f5b2ee12a89d224dc288ed28d9c3487caf72ed5bb69c585ca
  • code_bundle_ref: agn://artifact/17f927b54573758cd50d381ea871a004103ec03e6fd37eb1fa4115b6b4b6af35
  • archive_ref: agn://artifact/ac3522013a61af29e532eafed2967200e2242e8143a0aa510df36badbd49a76f
  • trace_index_ref: agn://artifact/8d0fce9af7d3306c27edd526c16bb804cb1dd07e28056d47d281a02e7b0db869
  • research_repo: /Users/macstudio_alexxon/Documents/AGN_SelfResearchProtocol