Data & Signal

Galaxy morphology — label-efficient transfer learning

17,736 real survey images · ResNet50 · Galaxy10-DECaLS

• I fine-tuned a ResNet50 on 17,736 real images from the Dark Energy Camera Legacy Survey, labelled by the Galaxy Zoo citizen-science campaign, across ten morphological classes.

  1 receipt

  • galaxy-morphology-transferMethod, splits, hyperparameters and full results.

• I chose macro-F1 as the primary metric rather than accuracy, because the classes are imbalanced and accuracy would have flattered the model on the common ones. I trained with class-weighted loss and stopped on validation macro-F1.

  1 receipt

  • READMEMetric choice stated with the reason; AdamW, ReduceLROnPlateau, early stopping.

• Then I ran the same experiment at 1%, 5%, 10% and 100% of the labels, evaluating all four on the identical held-out test set — because the question I actually wanted answered was not "how good can this get" but "how few labels do we need to pay for".

  1 receipt

  • READMEStratified subsets, one shared test set, four measured points.

• The most interesting result was an error analysis, not a score: the smallest class scored perfect recall while a mid-sized one scored worst. Distinctiveness beats frequency — so annotation budget should go to the ambiguous classes, not the rare ones.

  1 receipt

  • READMEPer-class recall — Cigar-shaped Smooth 100.00% on 334 training examples; Disturbed 59.26%.

The findingFive per cent of the labels — about 709 images, one afternoon of expert work — gets you to 75.3% accuracy. Ten per cent gets you 78.1%. That second doubling costs the same again and buys under three points. If I were advising someone paying for annotation, that is the entire recommendation, and it is worth more than the 87.3% I got from labelling everything.

Honest note: the training code for this one is not published yet — the repository currently holds the write-up and the results. It is on my machine and it is going up.

Two detection pipelines, and one I am not showing you

Simulation study — signal injection and recovery

I wanted to know whether a method could find a thing it had not been told the location of. So for each of three problems I generated a catalogue, planted a known signal inside it, and then measured how much of that signal the detector got back. Two of the three are below. The third fitted a value for dark energy that is not physically possible, which means the pipeline is wrong somewhere I have not found yet — so it is not on this page, and that is the only honest place for it to be. None of these use real observational data, and I would not present them as if they did — the whole point of injection-recovery is that you know the right answer in advance and can therefore grade the method honestly.