I am pleased that you want to use cryptocurrency mining to solve important scientific problems. I have briefly looked over your draft. Keep in mind that people in the past have tried to come up with useful proof of work problems with varying degrees of success, and the idea of using proof of work to solve problems in machine learning is nothing new; in fact, if you ask a random person who knows a little bit about cryptocurrencies to come up with a useful proof of work problem, then I would suspect that such an entity would probably use proof of work to solve problems in machine learning or something similar. I do not have any interest of being another coauthor for that paper, but I am willing to give my thoughts on it, and you can instead give me an acknowledgement.
What I like.It is good that you have not only realized that cryptocurrency mining algorithms need to be useful but you have also correctly identified Primecoin as a cryptocurrency with a useful proof of work problem but that Primecoin mining is of limited use. I have a Ph.D. in mathematics, and I have never heard of Cunningham chains before Primecoin, and I only see a little bit of merit in computing those Cunningham chains. It is also good that you have correctly recognized that training neural networks is more important than finding random chains of prime numbers. Plenty of mathematicians propose to use cryptocurrency mining for solving very specific mathematics problems that have very limited use. If these mathematicians thought these proposals all the way through, then they would realize that it is not a good idea to have a useful proof of work problem with questionable use because that will just annoy people.
The proposal that you have satisfies some of the requirements for proof of work problems. For example, in your proposal, it is good that the optimization problem for the current block is generated by the hash of the state of the blockchain so that the solutions to the proof of work problem cannot be reused elsewhere.
This proposal was quite readable and understandable.
My concerns.Your sort of proof of work problem may be good for a centralized cryptocurrency where a centralized authority is in charge of selecting and validating all of the problems (this defeats the purpose of having a cryptocurrency), but your proof of work should not be used as a replacement for Bitcoin's mining algorithm. A cryptocurrency mining algorithm where users post problems to solve has many problems and such a mining algorithm cannot be a replacement for SHA-256 mining. SHA-256 has many cryptographic properties that enable it to be used as a mining algorithm. What we need to use is a mining algorithm that has those cryptographic properties that we want. Here are some issues that I have with your proposal (some of these issues apply to most 'useful' proof of work problems; please forgive me or clarify if I did not interpret a part of your proposal correctly):
0. I would appreciate it if you left your full name along with the name of the coauthor of the paper, so that we can know who you are. You are not even an active user on this site, so we currently have no way of knowing anything about you.
1. Cryptocurrency mining algorithms must be quickly verified by everyone in the network. How do you expect everyone to quickly verify that the solutions to the problems posted on the blockchain? Furthermore, if you want to post all of the weights and biases of the neural network onto the blockchain, then this will take a lot of space.
2. Cryptocurrency mining algorithms must be progress free. This means that a miner who started mining 5 minutes ago but who has not achieved the block will have no advantage over a miner who started mining 2 seconds ago. This is clearly not the case with neural networks. If a mining algorithm is not progress free, then a user with a $10000 mining machine will win much more than 10 times the amount of blocks than a user with a $1000 mining machine. This will cause centralization because a few users will win all the blocks.
3. How long are the blocks for this cryptocurrency? If they are 2 minutes long, then is 2 minutes enough time to train a neural network?
4. The cryptocurrency mining algorithm also needs to be optimization free. How do we know that someone cannot discover a new trick that applies to your specific sort of classification problem that allows them to mine 100 times as quickly?
5. There is already a market for optimizing neural network. It is better to construct a useful proof of work problem for creating new markets that will be important in the future rather than just supplementing existing markets.
6. Any secret algorithmic innovation that anyone makes in computing your neural networks faster will be a source of insecurity of the blockchain since the algorithm innovators will have an advantage over those who did not innovate yet. In any case, your blockchain will likely lose because if there is no algorithmic innovation, then the proof of work is not as useful as you would like it to be, but if there is an algorithmic innovation, then the blockchain would have a security weakness.
You may be able to tweak your proposal so that it satisfies some of the properties that we would like to satisfy, but I doubt that one can tweak a proposal like this so that is about as secure as Bitcoin's mining algorithm while having some scientific merit.
A completely different proposal.I have launched a cryptocurrency called Circcash with a useful proof-of-work problem. Circcash's mining algorithm called Hashspin is designed to accelerate the development of the reversible computer. Can you spot any weaknesses in an algorithm like Hashspin? If you look at what Hashspin does, you will see that Hashspin satisfies all the security properties that cryptographic hash functions satisfy. I am not yet confident that Hashspin is more secure than SHA-256, but with a little bit more research, we can be confident that Hashspin (or a mining algorithm like Hashspin) is more secure than any well-known cryptographic hash function.
https://github.com/jvanname/Zammazazzer/blob/master/CirclefishICO.pdfIf you can find any weaknesses with Hashspin or Circcash (other than those stemming from its obscurity), then let me know, but if you think that Circcash is what we want in a useful proof of work problem, then you should mine Circcash and support the network.
