I’d like to know if there are around the AE community projects and teams that work on the following topics:
Machine learning (ML) as a service on blockchain. It combines PoW machines (having GPU) with ML data processing. Owners of GPU machines can unite/connect to clusters as it is going with PoW now (but for good things :). Distributed ML processing and data storage.
Federated and anonymous ML data processing.
Fully autonomous (ideally without human intervention, smart cities, supply chain etc) IoT economy ruled by data and ML.
Data market with ML processing as a service.
Stable coins collateralized by financial (AE, BTC, ETH) and nonfinancial assets (Services, ML processing, data storage).
Prediction markets mix of Kaggle and Augur, where ML models (not humans) make prediction on the event and solve ML tasks.
ML based anomalies, hacks, and fraud detections on the chain.
In your opinion, does aforementioned development direction have some value for AE?
Does AE blockchain technically ready for this task (data tx speed and so on)?
Or maybe community see another niche for AE just for finance and money transfers and payments.
Some interesting ideas, but I think AE is pretty happy with cuockoo pow
. That may change later on, but for this year the roadmap is already packed. I think AI or ML could interact with the chain. Maybe it could be used for answering oracles, i don’t know if this could even be prevenented.
There is nothing wrong with cuockoo PoW. I don’t offer to replace it with PoML (whatever it means). I even tried to mine some AE two days on my computer with GPU (without success though :). But this is not the point. As a ML practitioner I know there is a demand for the cheap GPU processing power. And people could give out their game computers for this task for some AE tokens. That is the idea.
I think that every blockchain should have the main sounding idea behind it that enrich it. And frankly, I don’t see what differs AE from other project. The same smart contracts, the same Oracles, the same everything. Erlang and functional paradigm are cool , but who cares about technology under the hood? What for can I use this blockchain, that is the point.
I have difficulties explaining it but discussions around making PoW “useful” instead of just solving a puzzle has been around for a pretty long time. But because of some features PoW brings (i think it is the difficulty that gets adjusted based on mining power) it is not trivial to develop a PoW algorithm that does something useful instead of just puzzles.
BUT John Tromp once said this to a reporters question:
How is your algorithm solving the problem of mining distribution with ASIC-miners involved?
Distributing (i.e. decentralizing) mining is a very tough, if not
unsolvable, problem that is mostly independent of the choice of PoW
algorithm. It requires a very large daily issuance in order to
incentivize multiple ASIC manufacturers to produce increasingly
efficient miners, and over the course of decades make mining equipment
more commoditized. Cuckoo Cycle, although conceptually quite simple,
is admittedly worse for in this regards than Bitcoin’s
SHA256-hashcash, as it will take longer to converge on the optimal
memory design. But in return, such designs are more likely to benefit
computing at large than are optimal SHA256 circuits.
He believes that with ASIC mining hardware build for Chuckoo Cycle we will get hardware that will benefit computing at large in the long term.
I’ll ping @ssh, maybe he has an better answer to this.
I don’t think @lostinio was suggesting to use machine learning tasks for PoW but just saying that it would be nice to have a market for rentable GPU power.
As for the other point, if securing the chain is not useful work, then we should re-evaluate what we are doing here (There’s a bit more to this but I don’t think that’s the purpose of this thread)
I beg your pardon for naming ML work “good things” as opposed to PoW, it was a joke. Certainly, PoW and security of the chain is the main task, and PoW is not useless waste of energy. Maybe this phrase distorted my idea.
I am not offering to replace PoW with PoMLW (let’s name it prove of machine learning work :). I thought about it earlier, and read what other guys think about it, and finally came to conclusion that at the moment it is not possible. These tasks are almost orthogonal each other in some way. I thought about wrapping (or mapping) ML tasks into pure cryptographical functions. But couldn’t find any concrete implementation. In ML, the main job is to find the extremum, usually minimum of the loss function, and it requires much power and memory. I thought in that direction, how can I wrap this task into math crypto domain. But ML processing is not a pure function, so we should wrap them into Monads and everything starts to be unpredictably complicated. It is very interesting task but requires many smart guys and years to solve it. Or maybe find another way how to come to consensus without pure functions.
I just offer to let people to share their idle processing power, usually game computers with GPU. And get AE tokens for it. We have many open source solution for distributed ML. Kubernetes with Dockers, Horovod and so on. To make your PC or server the part of the distributed ML you just need to download the Docker image and start it. Almost the same as to take part into mining process. In this way everyone can choose what to do: ML processing or mining.
ML is not replacement of the mining, and even not addition to it. ML processing is just additional value for blockchain infrastructure.
Moreover. The part of this distributed processing power can be used as a service on the AE blockchain to build value added services over it. For example: IoT Data market with data processing, prediction market kind of Augur on steroids :), anomaly/fraud detection on the blockchain, there are many of them. So any teams that want to build something on the AE get such additional value as ML, and it can be built into smart contract language as API.
In addition to my theoretical thoughts regarding PoMLW as a substitution of crypto PoW. The main principle of crypto PoW is - “Hard to find solution and easy to check it with a uniquely repeatable result in any point of space and time”. Therefore the question is - “Can we find a kind of isomorphism between this and PoMLW?” At first glance it seems yes. Training in ML is usually hard, but inference is usually fast. And that’s all, the result cannot be guaranteed. Maybe we can find the class of ML functions that satisfy these requirements (probably combinatorial tasks) but these class is very restricted (searching DNA sequence) and can be applied just to narrow class of ML problems.
(There’s a bit more to this but I don’t think that’s the purpose of this thread)