Your work

Your questions, projects, research drafts, saved investigations and ideas to review.

Kept in this browser

Saved in this browser. Clearing site data removes these records.

Checking this browser…

Keep a copy

Checking this browser’s storage protection…

This browser may clear saved work after a period without visits. Safari may do so after seven days of browser use without interaction here. Download a copy to keep it.

The copy is one file. Restoring it, in this browser or another, adds back only what is not kept there already; nothing kept there is replaced.

Restore from file

Recoveries (0)

Earlier saved text that this version could not read is kept here, inside recovery records. Recoveries are included in “Download all my work”. The store holds up to 64 records and 4 MB; no recovery is discarded automatically.

No recoveries are kept in this browser.

One real model failure

Is this bird classifier ready to ship?

It is right 97% of the time on photos mixed like its training data. Ask about it below: a scripted partner answers, with lines written in advance.

Work it through

landbirds on land99.6% · 225 of 226 right

landbirds on water69.3% · 165 of 238 right

waterbirds on land53.6% · 37 of 69 right

waterbirds on water95.5% · 64 of 67 right

600 validation photos, one dot each: filled when the classifier is right. No photographs are shown.

Data: the Waterbirds benchmark (Sagawa et al., 2020, Distributionally Robust Neural Networks for Group Shifts, ICLR), built from CUB-200-2011 bird photographs (Wah et al., 2011) and Places scenes (Zhou et al., 2017). The last-layer refit tested here comes from Kirichenko, Izmailov & Wilson (2023, Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations, ICLR).

Or bring your own question

About any model, paper or result you are working on.