Mining & models

NEB's proof of work is a training step of a small neural network. Mining the chain trains two models, and the chain stores their weights.

Difficulty
43.21 M
bits 0xbbc6c783
Expected attempts per block
45.31 T
2^256 / (target + 1)
Attempts / s (1 h)
758.94 G
network estimate from block times
Avg block time (1 h)
1 min 2 s
ASERT targets 60 s
Shares per block
21.1
average of the last 100 blocks
Next block trains
slot 0
kepler-v2, height 3,046

How it works, plainly

In Bitcoin, an attempt at a block is one hash. In NEB, an attempt is one gradient computation: the miner takes the block header, uses its hash to pick 16 samples from a public dataset, runs them through the current model and computes how each weight should move to make the model better on those samples. The gradient is hashed with the header into a ticket. If the ticket is below the target, the block is found. One attempt costs about 2 ms of a CPU core.

Attempts that miss the block target but meet an easier one (32 times easier) are shares. The next block lists up to 256 of them, and their gradients, summed, are applied to the model as one training step. So the work spent searching for blocks is not thrown away: it trains the model, and every node can replay the training from the genesis block and gets the same weights, bit for bit (integer arithmetic only).

The chain holds two models, one per dataset, and blocks alternate between them: a block at an even height mines on slot 0, an odd height on slot 1. Both are the same small convolutional network (KEP1, 26,000 weights of 4 bits each, between −8 and 7). Difficulty follows ASERT: it rises when blocks come faster than one a minute and halves every 10 minutes of delay.

Slot 0: Kepler transit signals

Light curves of stars watched by the Kepler space telescope, each folded around a periodic dip in brightness. The model learns to tell a planet candidate from a false positive (13,404 labelled signals).

Dataset
kepler-v2
Weights hash
6fb0997d29…ad0c8bd6
State at
block 3,045
Weights
26,000 (4-bit, 13,000 bytes packed)
Changed since genesis
3,199 of 26,000 (12.3%)

Layer 5 weights (32 × 768)

Hover or tap the map to read a weight.

Slot 1: TESS eclipse signals

Signals from the TESS mission (Astronet-Triage-v2, Tey et al. 2023), in the same two views. The model learns to tell a real eclipse from junk (2,363 eclipses, 4,726 junk signals).

Dataset
tess
Weights hash
a2e56dc60b…c973ee60
State at
block 3,045
Weights
26,000 (4-bit, 13,000 bytes packed)
Changed since genesis
2,844 of 26,000 (10.9%)

Layer 5 weights (32 × 768)

Hover or tap the map to read a weight.

The combined state hash of both models at the tip is f115829435…8dba8429. Data from the node's getmodel and getinfo.

Miner locks of the last 1,440 blocks (about a day)

By the lock of the first coinbase output. A pool pays from one lock, so one row can be many miners.

Specification: spec/shares.md and spec/nebhash-v2.md of the NEB repository. Block 3,045 difficulty 43.21 M.