Skip to content

T22: Build the XAI / shortcut-learning analysis and figure #337

Description

@mateuszwalo

Track: research · Level: core · Effort: ~8h · Depends on: #335 (T20)

Why this matters

Accuracy alone does not prove BNNR fixes the right problem. The XAI metrics and overlays show, visually and quantitatively, that ICD makes the model rely less on background shortcuts. This is the evidence that turns 'our numbers went up' into 'and here is why', which is what convinces a technical audience and a paper reviewer.

Steps

  1. From the per-run XAI metrics already produced by the benchmark (edge_ratio, coverage, gini), build a cross-condition table.
  2. Compute the correlation (Spearman) between edge_ratio and accuracy across runs.
  3. Assemble the OptiCAM overlay figure on the fixed validation indices, showing the model before and after ICD.
  4. Write the interpretation in findings.md: what the metrics say about shortcut reliance.

Done when

An XAI metrics table, a correlation result, and the overlay figure, all regenerable from the committed runs.


How to take this: comment "taking this" and wait to be assigned. Branch t22-short-desc from upstream/main, and put Closes #<this issue number> in your PR. Full workflow: the Cohort Handbook (pinned in Discord).

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions