A brief analysis of the decentralized computing power protocol Gensyn to speed up AI model training

Original source: TokenMore
Project Overview
Gensyn is a blockchain-based decentralized deep learning computing protocol that aims to establish an artificial intelligence (AGI) computing power market. It breaks down complex machine learning tasks into multiple subtasks and uses the computing resources of participants to achieve highly parallel computing. By automating task allocation, verification, and rewards through smart contracts, Gensyn eliminates centralized management and provides an efficient and autonomous solution for machine learning computing.
Official website: https://www.gensyn.ai/
Twitter: https://twitter.com/gensynai
technical background
With the rapid development of artificial intelligence technology, deep learning models have become more and more complex, and computing requirements have increased dramatically, but the available computing resources are relatively scarce. In this context, we are faced with a series of challenges.
First, to ensure the accuracy of the calculations, the validity of deep learning calculations needs to be verified. However, each layer in a deep learning model depends on the output of the previous layer, which makes verification complicated. We need to find ways to ensure that each step is performed correctly, especially as the models become more and more complex.
Secondly, there are also some problems in building a computing market. How to balance supply and demand, how to properly match computing resources, and how to motivate participants to contribute computing time are all difficult problems that need to be solved. The traditional market model may no longer be applicable in the computing field, and new approaches need to be explored.
Privacy protection is also an important issue. As global privacy regulations are strengthened, protecting the privacy of user data becomes particularly important. When building and training models, how to balance data privacy and model performance is a complex challenge.
In addition, in order to meet computing needs, high degree of parallelization has become an inevitable trend. Modern deep learning models need to be trained in parallel on large-scale hardware clusters to cope with the ever-increasing computing needs. The advancement of parallelization technology provides us with some hope to solve the problem of insufficient computing resources.
In summary, the field of deep learning computing faces many challenges, involving verification, market, privacy, efficiency, etc. Solving these challenges will help promote the continued development of artificial intelligence technology.
product design
The Gensyn protocol is like an intelligent computing network that is specifically designed to handle deep learning tasks. It enables people who are willing to participate in tasks with their own computers to be rewarded, just like helping others complete tasks. This protocol does not require middlemen or legal enforcement, but automatically assigns tasks and pays rewards through specific procedures. However, ensuring that tasks in this network are actually completed is a complex problem. Since each task depends on the results of the previous task, verifying the completion of the task is not simple. This problem is solved by combining three key concepts into a more efficient solution, making task verification more reliable.
• Probabilistic Proofs of Learning: Leverage metadata from the gradient optimization process to construct certificates of work completion that can be quickly verified by rerunning certain stages.
• Graph-based positioning protocol: Using a multi-granular, graph-based positioning protocol and cross-validating consistent execution, verification work can be re-run and compared to ensure consistency, which is ultimately confirmed by the blockchain itself.
• Truebit-style incentive game: construct an incentive game through the mechanism of staking and slashing to ensure that every economically rational participant performs tasks honestly.
Participants
There are four main roles involved in the Gensyn system: submitter, solver, verifier, and whistleblower.
• Submitters: These are the end users of the system who provide tasks that require computation and pay for the work done.
• Solver: is the main worker of the system, performs model training and generates proofs that need to be checked by the Verifier.
• Verifier: Crucial in connecting the non-deterministic training process with deterministic linear computation, they copy part of the solver’s proof and compare it to an expected threshold.
• Whistleblowers: As a final safeguard, whistleblowers review the work of validators and raise questions when they find problems in the hope of receiving a bounty.
Application
Process: Submit task -> Analysis -> Training -> Proof generation -> Verification proof -> Graphics-based precise positioning -> Cont -> Contract arbitration -> Settlement
Cost and performance
As Ethereum moves from proof-of-work to proof-of-stake, many miners will lose mining revenue. This presents a huge opportunity for the Gensyn protocol, allowing these miners equipped with machine learning hardware to be rewarded for using useful processor cycles, rather than just calculating hashes in a proof-of-work system. By attracting these mining resources and other potential computing resources, the Gensyn protocol has a cost advantage, such as the equivalent computing cost of an NVIDIA V100 will be 80% cheaper than AWS on-demand computing.
The performance of the Gensyn protocol was evaluated through Python simulations. Taking a small MNIST image classification model as an example, it was tested on a 6-core Intel Core i7 processor. The protocol was compared to 3 other methods: running the model locally (without using any protocol), using a Truebit-like replication method (7 validators), and running the model on Ethereum. Although the code lacks production-level optimization, the results show that the Gensyn protocol increases the time overhead when training the model by about 46%, but compared to Truebit-style replication, the performance is improved by 1,350%, and compared to running the model on Ethereum, The performance improvement is up to 2,522,477%. This shows that the Gensyn protocol has significant advantages in model training.
Team/Partners/Financing
6 members on LinkedIn:
https://www.linkedin.com/search/results/people/?currentCompany=%5B%2254109371%22%5D&origin=COMPANY_PAGE_CANNED_SEARCH&sid=dD *
Partner
Financing
• In January 2021, it conducted a Pre-Seed round of financing, attracting investors such as 7percent Ventures, Entrepreneur First, Counterview Capital and Id4 Ventures, with a financing amount of US$1.1 million.
• In March 2022, it conducted a seed round of financing, led by Eden Block, with 11 investors including Galaxy Digital, Maven 11, Coinfund, Jsquare, Hypersphere, Zee Prime, etc. also participating in the investment, with a financing amount of US$6.5 million.
• On June 12, 2023, a round of A financing was conducted, led by a16z, with investors such as CoinFund, Canonical Crypto, Protocol Labs, and Eden Block also participating in the investment, with a financing amount of US$43 million. These funds will be used to accelerate the launch of the protocol, expand the staff team, and recruit more machine learning engineers.
Gensyn has received over $50 million in investments at various stages of financing.
Project Summary
In general, Gensyn is a blockchain-based decentralized computing power protocol that is committed to allocating and rewarding machine learning tasks through smart contracts to accelerate the training of AI models and reduce costs. However, the prospect of decentralized training of large models still faces challenges such as communication and privacy, and feasibility needs to be re-evaluated.
In the development of AI, although there is potential to train large models using idle computing power, small AI models are more convenient and efficient in deployment and management. In many application scenarios, small AI models are still a more practical choice, and their value should not be ignored when chasing the craze of large models. Therefore, the development path of AI should consider diverse model sizes and needs to achieve wider and more flexible applications.
This article is from a contribution and does not represent the views of BlockBeats