A Proof of Useful Work (PoUW) for Artificial Intelligence on the Blockchain

[AGD]

Based on Bitcoin a few developers created a new white paper in which - instead of our known Proof of Work (PoW) - they introduce a protocol, based on a Proof of Useful Work (PoUW), which appears to be an interesting aproach. It even might be possible to implement this protocol into Bitcoin as a Hard Fork, which I still don't know if even possible, but I leave it here for discussion.

Full paper here: https://arxiv.org/pdf/2001.09244v1.pdf

What do you think?

[odolvlobo]

Bitcoin's concept of PoW relies solely on valueless work. The design is such that any value derived from PoW results in an additional equivalent amount of valueless work being done. In other words, you can't reduce the waste of PoW by producing value. Any PoW system that reduces waste by producing value would have to be a complete departure from Bitcoin's PoW.

The reason for this conundrum is that miners are incentivized to increase their hash power until their costs approach their revenue. Thus, increasing their revenue by generating value simply causes miners to increase their hash power (and their waste) until their costs again approach their revenue.

I think the only way out is to produce a true public good, but that is not as easy as it might seem. It also introduces potential points of failure into the economics of mining, for example where the public good stops being a public good or the system that ensures that it is a public good undermines the incentives to mine.

[ABCbits]

This is not the first paper/research to make PoW more useful, but most of them share same problem which is, certain party must give the data used to perform the computation.
It's obvious there's centralization risks, but my other concern are costs to verify the trained ML model and distribution of data required to perform the computation.

But it's quite interesting to use PoW for machine learning, so i'll read it thoroughly later.

[Wind_FURY]

Based on Bitcoin a few developers created a new white paper in which - instead of our known Proof of Work (PoW) - they introduce a protocol, based on a Proof of Useful Work (PoUW), which appears to be an interesting aproach. It even might be possible to implement this protocol into Bitcoin as a Hard Fork, which I still don't know if even possible, but I leave it here for discussion.

I don't know. It's a POW change, but if it's not against a threat to Bitcoin's existence, then I believe the community WON'T reach consensus.

Full paper here: https://arxiv.org/pdf/2001.09244v1.pdf

What do you think?

Staking? Like Decred? No no no with Bitcoin's unfair distribution.

[AGD]

This is not the first paper/research to make PoW more useful, but most of them share same problem which is, certain party must give the data used to perform the computation.
It's obvious there's centralization risks, but my other concern are costs to verify the trained ML model and distribution of data required to perform the computation.

But it's quite interesting to use PoW for machine learning, so i'll read it thoroughly later.

True. Even the SETI@home project was kind of a useful distributed PoW, besides the Altcoins that were refered in that paper (i.e. Primecoin etc) There also might be yet unkown additional attack vectors due to staking and PoS and also the Evaluators, which pay the nodes for the best work.

[AGD]

Based on Bitcoin a few developers created a new white paper in which - instead of our known Proof of Work (PoW) - they introduce a protocol, based on a Proof of Useful Work (PoUW), which appears to be an interesting aproach. It even might be possible to implement this protocol into Bitcoin as a Hard Fork, which I still don't know if even possible, but I leave it here for discussion.

I don't know. It's a POW change, but if it's not against a threat to Bitcoin's existence, then I believe the community WON'T reach consensus.

Full paper here: https://arxiv.org/pdf/2001.09244v1.pdf

What do you think?

Staking? Like Decred? No no no with Bitcoin's unfair distribution.

Yeah, I don't think there will ever be a consensus in the Bitcoin community about this one, but it has at least the chance to be one of the better Shitcoins in existence.

[HeRetiK]

After seeing the occasional thoughtless newbie post about combining blockchain with machine learning (who cares about details as long as we put buzzwords together amirite?) it's nice to finally see something with a bit more substance.

What bothers me a bit is that the concept of PoUW seems to be more like a red herring, with the actual security coming from a PoS system trying to keep participants honest. That is on top of what odolvlobo posted regarding PoW requiring "useless" work for security.

Their treatment of Sybil Attacks also looks a bit hand-wavy to me. Disallowing nodes to pick their own tasks might work if you only have a limited number of bad actors; I have my doubts that this would be effective enough in a more serious scenario.

Either way I still need to dig a bit deeper into that paper. Thanks for sharing!

[KTChampions]

Based on Bitcoin a few developers created a new white paper in which - instead of our known Proof of Work (PoW) - they introduce a protocol, based on a Proof of Useful Work (PoUW), which appears to be an interesting aproach. It even might be possible to implement this protocol into Bitcoin as a Hard Fork, which I still don't know if even possible, but I leave it here for discussion.

Full paper here: https://arxiv.org/pdf/2001.09244v1.pdf

What do you think?

1.) It seems to me that such an improvement should not be considered relative to bitcoin. There is already a consensus here and network power is being spent on security. This is the greatest usefullness for the bitcoin network.

2.) As for creating a new currency, I like this idea. It turns out that useful work from the very beginning of mining will remove inflation from the coin. It will be a fairly simple and logical chain: work -> payment for it. This will hit decentralization and this is the main drawback, but such a coin should find its niche in the market. And the likelihood that this coin will be supported by governments is very high.

[FGT]

As you know Tesla Autopilot is the best one, but why?
The reason is their Machine Learning Network, it is huge. Another Car Manufactures don't have such networks, but the time is coming. In nearest feature many providers start working with integrating Artificial Intelligence and Autopilots.

So this idea of using mining machines for useful work is a real answer.

But actually I don't believe that it could happen in Bitcoin network, may be in some new Altcoins.

[PrimeNumber7]

The paper mentions that miners would switch to bitcoin mining if it becomes more profitable to mind bitcoin than to mine PoUW, however I don't believe bitcoin ASICs can perform this kind of work.

The paper also mentioned that a (rationally acting 'customer') would switch to Google Cloud, Azure, or AWS ML tools if the price became higher than the competition, however this sounds more complicated than any of these tools. Most cloud ML tools also can offer somewhat of a predictable cost.

It also appears the network is vulnerable to a sybil attack if a bad actor creates enough evil nodes/miners. if this were to become a legitimate threat to AWS, Amazon could deploy its servers to produce bogus models, and end users may not realize this to be the case until their decisions they base on the returned models turn out to be bad.

Testing the model with testing data can also be expensive for the 'evaluators' if they are needing to validate multiple models, and this testing may be in vein if miners are producing not poor models.