ModelsEvery NEB block is mined on a neural network and its shares train it. This is how well the networks have learned.

Accuracy history verified against our nodeThe NEB project publishes the accuracy of each model on samples it never trained on. At block #4,171 the models in that history have the same weights hash as the models our own node reports.
Updated 2m ago
KeplerKepler transit signals (planet candidate or false positive), even heights
Accuracy
89.7%
Since genesis
+39.3 pts
Last 24h
+0.7 pts
14,961 training steps in 24h, started at 50.4%. As of block #4,171.
false positiveplanet candidate
32 hidden neurons (brighter = stronger) linked to the outputs: supports or argues against that output.
Weights hash at block #4,172, from our noded19b580c50028cf5…82961ed2
TESSTESS signals (eclipse or junk), odd heights
Accuracy
94.0%
Since genesis
+79.9 pts
Last 24h
−0.4 pts
15,236 training steps in 24h, started at 14.1%. As of block #4,171.
junkeclipse
32 hidden neurons (brighter = stronger) linked to the outputs: supports or argues against that output.
Weights hash at block #4,172, from our node604ede5a2f80caa5…d3a5717d

Accuracy over time

Share of held-out samples each model classifies correctly, as it trains block by block
Kepler89.7%TESS94.0%

What the models learn

Kepler: Kepler transit signals (planet candidate or false positive), even heights. Its outputs are false positive and planet candidate.

TESS: TESS signals (eclipse or junk), odd heights. Its outputs are junk and eclipse.

A share is a mined attempt that also counts as one training step: the gradient of the attempt is applied to the model, so the more the network mines, the more the models learn. Accuracy is measured by the NEB project on samples kept out of the chain; we check that the models in that history are the models of our own node, and record the weights hash of every block.