The Reward Mechnism
An incentive mechanism in the DIN protocol encourages contributors to submit data to improve the model’s accuracy, the validator to validate the data's correctness, and the computation node to convert datasets to vectors.
The incentive mechanism combines gamification points (in the system named "Wafer") and a methodology based on prediction markets.
For the data contributors, participants are rewarded according to how well they improve the model’s performance as measured by the test data.
Reward for data contribution
There are three phases, as shown in Fig 1.
In the commitment phase, the provider deposits the bounty and defines a loss function L(h, D). This function measures the model's loss on any dataset D (typically the average loss on points in the dataset). Finally, the provider cryptographically commits to a test dataset, a small random fraction of which is initially revealed.
In the participation phase, people add data or otherwise update the model. Each participant must stake some units of network token along with their update. After an end condition is met (such as a maximum time limit or amount of data), this phase ends. A new cycle can begin if a new provider commits new test data.
In the reward phase, the provider uploads the test dataset, and the smart contract checks to ensure it satisfies the commitment. Then, the smart contract determines rewards for all participants, as discussed next.
Reward calculation
First, imagine that the bounty B = 1, so that each participant’s reward is their stake plus the following number:
This reward function is proposed here, based on automated-market-maker or scoring-rule-based prediction markets. It can be pictured as follows: The smart contract first pays L(h0, D) to the first participant. Their data updated the model to h1, so they spent the second participant L(h1, D). This continues until the last participant pays L(ht, D) back to the smart contract. The better ht performs, the fewer participants t has to pay forward, so they are incentivized to provide data as useful as possible relative to the (expected) test set. (If ht performs worse than the previous model, t loses some or all of their stake.) In total, the smart contract pays out a net amount of L(h0, D)−L(ht , D), which is the total improvement from all contributions. It is at most 1 by assumption on the loss function.
Finally, we must scale this mechanism for a bounty of B The approach proposed by J.D. Abernethy would require that all participants stake B, which is infeasible. Therefore, the approach of iterating the mechanism B times is used instead. In each iteration, the participant stakes some tokens and receives a reward. The participant will drop out if he can no longer stake tokens due to losses. Although this is slightly complex, it remains that the better ht, the more reward t gets, so we believe incentives for participation are strongly aligned.
The reward for data validation and computation
coming soon..
