A self-supervised map of thousands of real JWST galaxies.
A SimCLR encoder was shown thousands of JWST cutouts with no labels and learned to place galaxies that look alike near each other. Here is its 2-D map (UMAP). Drag to fly, scroll to zoom; thumbnails appear as you dive in. Step the colour key left to right: the three sky regions are mixed (location doesn't drive the map), but morphology separates, and at the far right anomaly lights up the weird ones. None of these labels were seen during training. loading atlas…
What you're seeing. Each point is a real JWST/NIRCam galaxy (GOODS-S, COSMOS, UDS), pulled from the DAWN JWST Archive. A ResNet-18 trained by self-supervised contrastive learning (SimCLR, NT-Xent) maps each cutout to a vector; UMAP flattens those vectors to this 2-D layout. Morphology labels from Galaxy Zoo only colour the map; the geometry is the encoder's alone. Colour by anomaly (far right) for the payoff: the known Little Red Dots appear automatically and land where the encoder flagged the unusual, with no labels.
Flip the colour from sky region to morphology: with no labels, the encoder has already sorted galaxies into smooth, featured and merger.
how it knowsSimCLR pulls two random crops of the same cutout together and pushes different galaxies apart, so look-alikes end up neighbours.
The Galaxy Zoo votes only colour the finished map; they never touched training.
Under sky region the three deep fields stay thoroughly mixed: a disk in GOODS-S sits beside one in COSMOS.
but not nothingA classifier can still guess the field 63% of the time (±0.5% across folds, vs 41% by chance), but that's mostly survey depth and PSF, not real structure.
And the Red-Dot signal isn't a depth artefact: retrain the encoder with a whole field held out, and the anomaly enrichment on that unseen field survives, so it's about galaxies, not the survey.
Colour by anomaly: 216 known Little Red Dots (a separate published catalogue laid over the atlas, not drawn from it) pile into the same corner the encoder flagged as weird, though it was never told they exist.
why they stand outLRDs are compact, uncommon and genuinely red, so they sit far from everything in latent space → a high anomaly score (7.4× enriched in the top 10%; 95% CI 6.8–7.9×, p<0.0001 against a random-galaxy null). The 216 are every known Little Red Dot in these fields, a fixed set.
The encoder is point-source aware: each cutout is centred on its source and stretched against the real sky noise, so empty sky stays dark instead of becoming colour static. That sharpened the signal from ~5× to 7.4×, and it holds up even on a deep field the encoder was never trained on.
The honest test: rank the atlas by anomaly, take the reddest, most compact outliers that aren't in any known catalogue, and check them. The pipeline rediscovered two published LRDs on its own but turned up no genuinely new one. It's a validated method, not a discovery claim, and we'd rather say so.
the full storyThe two rediscoveries (Akins+24 COSMOS-Web dots, matched to 0.14″) are the real win: the encoder flagged them with no labels, and they pass the published selection cuts: an end-to-end validation. The rest of the shortlist looked red only in quick-look photometry; measured in the full survey catalogues they come out at F277W−F444W ≈ 0, not the ≈1 of a real Red Dot.
The reason is built in: this atlas is seeded from bright, human-classified Galaxy Zoo galaxies, so it can't reach the faint population where undiscovered Red Dots hide. Catching those needs a fainter, JWST-native starting sample, the next build.
These galaxies live in three deep fields chosen to look out of our own Milky Way. Spin the globe: the three fields sit near the poles of the gold galactic plane (far from its dust) and at varied distances from the cyan ecliptic (Earth's orbital plane).
drag to rotate · gold = galactic plane · cyan = ecliptic · the three coloured dots = the deep fields (same colours as the atlas "sky region" mode) · the Sun's spot is for today
galactic latitude b is the angle each sightline makes with the disk; bigger means a cleaner view out. (schematic, not to scale)