BitTensor τ — democratizing ML

What is it about?

BitTensor plays with Bitcoin and Tensor words on its name, we can think about it as a democratize platform for machine learning.

Tensor derives from the operations that neural networks perform on multidimensional data arrays, which are referred to as tensors, and Bit(coin) is a reference to blockchain.

BitTensor aims to create a marketplace for neural networks, resistant to collusion (price manipulation), which main goal is to provide proper incentives to improve intelligence generated from information

Why we need something like BitTensor?

Data today is considered a commodity, and as in any other market, you have the problem of defining the price. Questions like “What is a fair price for this commodity?” or “How to avoid price manipulation?” needs to be answered in order for a marketplace to succeed.

Another problem of marketplaces is when few entities have access to the best sources (centralization) and, when talking about intelligence you can observe that everything that is not “state of the art” is discarded (exclusion), even if the intrinsic value produced is high.

BitTensor aims to solve these problems in the domain of data (more specifically intelligence derived by neural networks), using the combo of transparency + decentralization + incentives that blockchain provides… but how?

Researchers can monetize machine intelligence work and consumers can directly purchase it (with a fair price)

How does it actually work?

Before going into technical explanations we can say something like:

BitTensor proposes to use intelligence to measure intelligence

Sorry… what???

The idea is that we can use intelligence to discover the value of models available on BitTensor marketplace avoiding collusion, price manipulation and exclusion.

Think about BitTensor as a network of peers that share some intelligence, these intelligence is valued based on scores provided by each peer.

To avoid price manipulation BitTensor uses mathematical models to put in place proper incentives for good actors and make bad actors behavior not relevant/profitable (mathematically).

The mathematics behind are based on 4 concepts (peer-ranking, incentives, bonding and consensus):

• First is applied a peer-ranking, each peer takes outputs from other peers as inputs and generate a weight.
  Each peer has also a weight based on their stake.
  The combination of both determines the incentive paid to each peer.
• This will not solve collusion as peers would only vote for themselves, this is why the incentive formula is added.
  The idea is that peers get more weight if they have higher consensus with other peers, this avoids the self-voting problem as you need to rank other peers to reach consensus
  You can see an example graph using Sigmoid function where
  more consensus = more incentives:

• Now we solved naive collusion of small groups controlling the inflation, but still we need a mechanism to provide proper weights, for this we add bonds
  My understanding of bonds is that a peer A that bonds with another peer B will get higher incentives if that bonded peer B increases its value. It’s like betting into the best horse of a race, and this forces to chose wisely which peers you bond with (and hence promotes good quality peers to get higher weights as everybody will try to bet into the winning horse)
• The last part is to avoid that a lack of consensus where small sub-graphs get most of the incentives. Here BitTensor adds a loss factor applied to sub-networks that have less than 50% consensus, making their stake irrelevant after a set of iterations (incentivizing the majority network over the sub-graph)
  The idea is, if you have less than 50% of consensus you are penalized

Example of a sub-graph with 49% of the stake, we can see on the next simulation how their stake becomes irrelevant after ~90 blocks
(being rewards % represented by a number between 0 and 1):