Research

Architecture search spends most of its budget scoring candidates. Predictors make that cheap, but the predictors of the time each understood one search space, because they read a fixed recipe of blocks.

The paper rewrites any convolutional network as a computation graph of primitive operations, then learns that graph with contrastive learning on unlabelled architectures from several families. A small head, fine-tuned on a few scored networks, ranks a family the predictor was not trained on.

I am a co-author, not the lead author. It was written at Huawei's Canadian lab in Edmonton, with the University of Alberta and UBC.

• 0.917

  Spearman on NAS-Bench-201

• 0.892

  Spearman on NAS-Bench-301, where zero-cost proxies go negative

• 94.23%

  best NAS-Bench-101 architecture found in 700 queries, on average the 2nd best in the space

• 79.2%

  ImageNet top-1 for a MobileNetV3 found with the transferred predictor

Spearman figures are means over five fine-tunes, with a spread of about 0.01. Search and ImageNet numbers are the single results the paper reports.

Labelled change pairs are scarce and every new site looks different. The question was whether a frozen vision transformer gives usable change maps out of the box, and what it takes to run one honestly on mixed satellite pairs.

The first zero-shot runs returned nothing. Logging the score distribution showed why: confidence sat between about 0.3 and 0.6, and a threshold borrowed from convention, 0.7, erased every change. Thresholds now come from the spread each run actually produces, and that spread is logged every time.

One coarse tile lost small features. Three tile scales, a classical channel and a confidence weight between them took one scene from 2 detections to 187. A light head trained on 26 real pairs could not be shown to beat the frozen comparator, so it was not adopted. The negative result stays on the record.

• 185

  change polygons on a public coal mine, 671.47 ha

• 2 → 187

  detections on one scene after multi-scale tiling, a count and not an accuracy

• 356 s

  for that scene on a 4 GB GPU, 2.8 GB peak memory

• 30.2%

  mean pixel confidence, the honest mid band; 177 of 185 polygons sit between 30% and 50%

Precision and recall against labelled change polygons have not been measured yet. A 200-pair labelled set is planned for that.

Old parcel sheets hold decades of records but carry no coordinates, and an ink drawing looks nothing like a photograph. The open questions: does any matcher find enough true correspondences between the two, how do you notice a wrong alignment when in-sample residuals look fine, how do you recover, and how much transform can a handful of control points support?

The answer was two graphs and a gate. The map becomes a parcel graph and the satellite base a road graph. A dense matcher proposes a seed across zoom levels, and the seed is scored by leave-one-out error, never in-sample. If it fails, a search over cluster shapes takes over, and with few control points the transform is capped at affine.

One bug is worth keeping: matches were rescaled by the original image size while the matcher worked at a fixed internal size, which quietly threw every match away. A test now pins it.

• 6.15 m

  trusted leave-one-out error with 8 control points, one golden run

• 4 sheets

  stitched into a 13,061 by 8,368 pixel canvas before matching

• ~75 min

  for a cold job on one GPU; about 5 minutes to rematch after an operator moves a point

• 5 m

  the most the optimiser will move an anchor a person placed

Internal measurements on one run. The in-sample residual is lower and is never quoted.

Frozen SAM2 and CLIP models, used as they are, with rules around them where they fail. Prompted on vegetation imagery, SAM2 drew field shapes with a median IoU of 0.562 against a published baseline at 0.086. A graph segmentation beside it recovered 90% of field boundaries.

Labelling needed fixing. On an arid mine, pit benches matched the built-up class, so low-colour masks are now tested by chroma and sent to bare soil. Ranking change by confidence times area filled one summary with a single class, so one result per class is now reserved before the rest fill by score.

• 0.562

  median field IoU, 6.5 times the published baseline

• 0.901

  boundary recall from the graph segmentation

• 0.809

  overall accuracy for seven crop classes, kappa 0.755, spatial cross-validation on 136 fields

Crop accuracy comes from a separate classifier on season features, validated by holding out whole areas.