start
  • ๐Ÿ‘‹Welcome
  • ๐Ÿ“–Introduction
  • ๐Ÿ’กUse Cases
  • ๐Ÿง‘Personas
    • Film Production
    • Animation Studios
    • Game Developer
    • Industrial Design
    • Advertising
    • AI Image Generation / Text-to-Image
    • Speech-to-Text & Text-to-Speech
    • AI Video Enhancement & Processing
    • AI Object Detection & Image Analysis
    • Enterprise LLM API
    • Private Knowledge Base LLM (RAG - Retrieval-Augmented Generation)
    • Family Photographer
    • Indie Game Developer
    • Aspiring 3D Artist
    • Playstation Gamer
  • ๐Ÿš€Get Started
    • Janction Node Operation Graphic Tutorial
  • ๐Ÿ”—Related Work
  • ๐Ÿ—๏ธArchitecture
    • Actor Model
  • ๐Ÿ–ฅ๏ธPooling
  • ๐Ÿช™Token
  • โšกColocation of Idle Processor Computing Power
  • โœ…Proof of Contribution
  • ๐ŸŽฎGPU Marketplace
    • Pricing strategy based on pvcg
  • โ“HELP FAQ
    • FAQ
      • How Janction Efficiently Stores AI/ML Models for Different Users๏ผŸ
      • Compared to traditional cloud GPU platforms, how does Janction's distributed idle GPU computing powe
      • How does Janction ensure the efficiency and quality of data annotation for various data types with d
      • How does Janction's execution layer handle the various AI subdomain functionalities?
      • How does Janction select and use different DAs?
      • Is Janction considering adopting the security guarantees provided by Restaking?
      • What is the current progress of Janctionโ€™s product technology?
      • How will Janction consider airdropping to the community?
  • ๐Ÿ›ฃ๏ธRoadmap
  • ๐Ÿ“œPolicy
    • Terms
Powered by GitBook
On this page
  • Crรฉmer-McLean Mechanism
  • Procurement-VCG (PVCG) Mechanism
  1. GPU Marketplace

Pricing strategy based on pvcg

PreviousGPU MarketplaceNextHELP FAQ

Last updated 11 months ago

In the role benefit distribution model of distributed machine learning, the benefits of the global verifier and local verifier are almost fixed like the miners in the blockchain, and will only change with the activity of the system. The aggregator is the task scheduler and coordinator, and its benefits are positively correlated with the trainer. The only more complicated issue is that during the local training process of the model, the trainer may use data from other data sources, and we need to give additional consideration to the gaming problem in the local training scenario of distributed machine learning. The trainer's operating environment will have some thresholds, while other network participants who do not fulfill the hardware requirements can participate in the training process by providing data. A game model is needed between the data providers and the real trainers to distribute their benefits fairly.

To incentivize data providers to contribute the best datasets to the training process, we need to pay sufficient rewards to data providers to cover their costs. The marginal monetary reward for contributing more data should be no less than the resulting marginal cost. In addition, our goal is to maintain a balanced budget and optimize social welfare. At least three sources of information asymmetry are intertwined in this problem: 1) The dataset owned by each data provider; 2) The cost of each data provider; 3) The valuation of trained distributed machine learning models by model users.

To overcome these information asymmetries and achieve the above objectives, we design rational incentives, i.e., a function that calculates participants' payoffs.

There exists a set of nnn data providers, denoted by N=0,โ€ฆ,nโˆ’1N = {0, \ldots, n - 1}N=0,โ€ฆ,nโˆ’1 , and another set of mmm model users, denoted by M=n,โ€ฆ,n+mโˆ’1M = {n, \ldots, n + m - 1}M=n,โ€ฆ,n+mโˆ’1. Each data provider iโˆˆNi \in N iโˆˆN owns a dataset dห‰i\bar{d}_i dห‰iโ€‹. It claims it owns a dataset d^i\hat{d}_i d^iโ€‹. The federation accepts a dataset diโŠ˜d^id_i \oslash \hat{d}_i diโ€‹โŠ˜d^iโ€‹ from this data provider. We call ฮทi=did^i\eta_i = \frac{d_i}{\hat{d}i} ฮทiโ€‹=d^idiโ€‹โ€‹ the acceptance ratio, where โŠ˜\oslashโŠ˜ denotes element-wise division. Trained on datasets d=(d0,โ€ฆ,dnโˆ’1)d = (d_0, \ldots, d{n-1})d=(d0โ€‹,โ€ฆ,dnโˆ’1) from all data providers, the usefulness of the distributed machine model is Q(d)Q(d) Q(d). Model users may be granted limited access to the distributed machine model such that the usefulness of the distributed machine model to model user jjj is ฮบjQ(d)\kappa_j Q(d)ฮบjโ€‹Q(d), where ฮบj\kappa_j ฮบjโ€‹ is called the access permission. Each data provider iโˆˆN i \in NiโˆˆNhas a cost type ฮณiโˆˆฮ“i\gamma_i \in \Gamma_i ฮณiโ€‹โˆˆฮ“iโ€‹. Its cost of contributing data did_i diโ€‹ is c(di,ฮณi)c(d_i, \gamma_i)c(diโ€‹,ฮณiโ€‹). The collection of cost types of all data providers forms the cost type profile ฮณ=(ฮณ0,โ€ฆ,ฮณn)\gamma = (\gamma_0, \ldots, \gamma_n)ฮณ=(ฮณ0โ€‹,โ€ฆ,ฮณnโ€‹). Data provider iii may report a different cost type ฮณ^i\hat{\gamma}_iฮณ^โ€‹iโ€‹.

Each model user jโˆˆMj \in MjโˆˆM has a valuation type ฮธjโˆˆฮ˜j\theta_j \in \Theta_j ฮธjโ€‹โˆˆฮ˜jโ€‹. Its valuation on the trained distributed machine model is

w(ฮบjQ(d),ฮธj)=v(d,ฮบj,ฮธj)w(\kappa_j Q(d), \theta_j) = v(d, \kappa_j, \theta_j)w(ฮบjโ€‹Q(d),ฮธjโ€‹)=v(d,ฮบjโ€‹,ฮธjโ€‹)

The collection of valuation types of all model users forms the valuation type profile ฮธ=(ฮธn,โ€ฆ,ฮธn+mโˆ’1)\theta = (\theta_n, \ldots, \theta_{n+m-1})ฮธ=(ฮธnโ€‹,โ€ฆ,ฮธn+mโˆ’1โ€‹). Model user jjj may report a different valuation type ฮธ^j\hat{\theta}j ฮธ^j. The payment to data provider iโˆˆNi \in N iโˆˆN is piโ‰ฅ0p_i \geq 0 piโ€‹โ‰ฅ0. The payment to model user jโˆˆMj \in MjโˆˆM is pjโ‰ค0p_j \leq 0pjโ€‹โ‰ค0. We denote ps=(p0,โ€ฆ,pnโˆ’1)p_s = (p_0, \ldots, p{n-1}) psโ€‹=(p0โ€‹,โ€ฆ,pnโˆ’1) and pd=(pn,โ€ฆ,pn+mโˆ’1)p_d = (p_n, \ldots, p_{n+m-1})pdโ€‹=(pnโ€‹,โ€ฆ,pn+mโˆ’1โ€‹). The federation income is I=โˆ’โˆ‘j=nn+mโˆ’1pjI = - \sum_{j=n}^{n+m-1} p_jI=โˆ’โˆ‘j=nn+mโˆ’1โ€‹pjโ€‹; the federation expenditure is E=โˆ‘i=0nโˆ’1piE = \sum_{i=0}^{n-1} p_iE=โˆ‘i=0nโˆ’1โ€‹piโ€‹; the federation profit is P=โˆ‘l=0n+mโˆ’1plP = \sum_{l=0}^{n+m-1} p_l P=โˆ‘l=0n+mโˆ’1โ€‹plโ€‹. Participants' preferences are represented by quasi-linear utility functions:

ui(โ‹…)=pi(โ‹…)โˆ’ci(โ‹…)ย forย iโˆˆN;uj(โ‹…)=pj(โ‹…)+vj(โ‹…)ย forย jโˆˆMu_i(\cdot) = p_i(\cdot) - c_i(\cdot) \text{ for } i \in N; \\ u_j(\cdot) = p_j(\cdot) + v_j(\cdot) \text{ for } j \in Muiโ€‹(โ‹…)=piโ€‹(โ‹…)โˆ’ciโ€‹(โ‹…)ย forย iโˆˆN;ujโ€‹(โ‹…)=pjโ€‹(โ‹…)+vjโ€‹(โ‹…)ย forย jโˆˆM

The social effect of distributed machine learning is measured by social surplus, defined as

S(โ‹…)=โˆ‘j=nn+mโˆ’1vj(โ‹…)โˆ’โˆ‘i=0nโˆ’1ci(โ‹…)S(\cdot) = \sum_{j=n}^{n+m-1} v_j(\cdot) - \sum_{i=0}^{n-1} c_i(\cdot)S(โ‹…)=j=nโˆ‘n+mโˆ’1โ€‹vjโ€‹(โ‹…)โˆ’i=0โˆ‘nโˆ’1โ€‹ciโ€‹(โ‹…)

We set the data providers as the supply side and the trainers as the demand side, and use the PVCG model, which maximizes social welfare, as the supply-side auction offer model, and use the Cremer-McLean mechanism to maximize the demand-side utility.

Crรฉmer-McLean Mechanism

The detailed process of the Crรฉmer-McLean mechanism for the demand side involves the following steps:

Step 1: Consumer Valuation and Preferences

Step 2: Decision Rule

Step 3: Payment Rule Design

Crรฉmer-McLean Theorem Formulation

Optimization Process

Procurement-VCG (PVCG) Mechanism

As a counterpart of the Crรฉmer-McLean mechanism, we create the PVCG on the supply side. The PVCG mechanism is designed to incentivize distributed machine learning participants to truthfully report their type parameters and offer their best datasets to the federation. This mechanism provides theoretical guarantees for incentive compatibility, allocative efficiency, individual rationality, and weak budget balancedness. The PVCG mechanism, along with the Crรฉmer-McLean mechanism, aims to address the challenges of information asymmetry and free-riding in distributed machine learning by providing appropriate incentives to participants.

Procurement Auction Process of PVCG and Payment Calculation for Data Providers

Step 1: Data Providers Claim Datasets to Offer and Bid on Cost Types

As the first step, each data provider submits a sealed bid for their claimed datasets and cost types. The claimed dataset ( \hat{d}_i ) is the dataset that data provider ( i ) claims to offer for distributed machine learning. It may differ from the actual dataset ( \bar{d}_i ) owned by the data provider. Similarly, the reported cost type ( \hat{\gamma}_i ) may differ from the true cost type ( \gamma_i ).

Step 2: The Coordinator Chooses the Optimal Acceptance Ratios

Step 3: Data Providers Contribute Accepted Datasets to Distributed Lachine Learning

Step 4: The Coordinator Makes Transfer Payments to Data Providers According to the PVCG Sharing Rule

which includes consumer surplus Sd=โˆ‘j=nn+mโˆ’1vj(โ‹…)S_d = \sum_{j=n}^{n+m-1} v_j(\cdot)Sdโ€‹=โˆ‘j=nn+mโˆ’1โ€‹vjโ€‹(โ‹…) and producer surplus Sp=โˆ’โˆ‘i=0nโˆ’1ci(โ‹…)S_p = - \sum_{i=0}^{n-1} c_i(\cdot) Spโ€‹=โˆ’โˆ‘i=0nโˆ’1โ€‹ciโ€‹(โ‹…). There are user-defined unfairness functions wห‰s(ps,c)\bar{w}_s(p_s, c) wห‰sโ€‹(psโ€‹,c) and wห‰d(ps,v)\bar{w}_d(p_s, v) wห‰dโ€‹(psโ€‹,v) that measure the unfairness among data providers and model users.

Given that the federation Income I(Q)I(Q) I(Q) and the model quality Q(d^,ฮท)Q(\hat{d}, \eta)Q(d^,ฮท) are exogenous functions, the supply-side distributed machine learning incentive mechanism design is to design the optimal. Specifically on the supply side, letting the data providers provide the maximum data efficiency ฮทi(d^,ฮณ^)\eta_i(\hat{d}, \hat{\gamma}) ฮทiโ€‹(d^,ฮณ^โ€‹) and provide the optimal reward pi(d^,ฮณ^)p_i(\hat{d}, \hat{\gamma}) piโ€‹(d^,ฮณ^โ€‹) for them, and on the demand side, letting the trainers produce the optimal model results so that the trainer outputs the maximum model validity ฮบj(ฮธ^)\kappa_j(\hat{\theta})ฮบjโ€‹(ฮธ^) and provides the optimal reward pj(ฮธ^)p_j(\hat{\theta})pjโ€‹(ฮธ^) for them.

Each consumer iii privately holds a valuation ฮธi\theta_iฮธiโ€‹ for the product or service being offered.

Consumers submit their valuations through a decision rule ฮบ(ฮธ^)\kappa(\hat{\theta}) ฮบ(ฮธ^), where ฮธ^\hat{\theta} ฮธ^ represents the reported valuations.

An interim incentive compatible and interim individually rational payment rule p(ฮธ^)p(\hat{\theta})p(ฮธ^) is designed to extract the full consumer surplus.

The Crรฉmer-McLean Theorem states that for any decision rule ฮบ(ฮธ^)\kappa(\hat{\theta})ฮบ(ฮธ^) satisfying the Crรฉmer-McLean condition and identifiability condition, there exists a payment rule p(ฮธ^)p(\hat{\theta})p(ฮธ^) that extracts full consumer surplus:

โˆ‘j=nn+mโˆ’1pj(ฮธ^)=โˆ‘j=nn+mโˆ’1w(ฮบj(ฮธ^)Q,ฮธj)\sum_{j=n}^{n+m-1} p_j(\hat{\theta}) = \sum_{j=n}^{n+m-1} w(\kappa_j(\hat{\theta})Q, \theta_j) j=nโˆ‘n+mโˆ’1โ€‹pjโ€‹(ฮธ^)=j=nโˆ‘n+mโˆ’1โ€‹w(ฮบjโ€‹(ฮธ^)Q,ฮธjโ€‹)

The payment rule p(ฮธ^)p(\hat{\theta})p(ฮธ^) can be found by minimizing a loss function to ensure interim incentive compatibility, individual rationality, and full consumer surplus extraction.

When designing the supply-side mechanism, we assume the federation income I(Q)I(Q) I(Q) to be an exogenous function that depends on the quality of the distributed machine model QQQ. This assumption allows us to focus on optimizing the supply-side distributed machine learning incentive mechanism without directly considering the intricacies of how the federation income is determined. By assuming I(Q)I(Q)I(Q) as an exogenous parameter, we can streamline the design process and concentrate on factors such as dataset contributions, cost types, and acceptance ratios to achieve the desired objectives in distributed machine learning. The supply-side use of the PVCG mechanism gives us the revenue of the distributed machine learning training process as:

I(Q)=โˆ’โˆ‘j=nn+mโˆ’1pj(ฮธ)=โˆ‘j=nn+mโˆ’1w(ฮบj(ฮธ)Q,ฮธj)I(Q) = - \sum_{j=n}^{n+m-1} p_j(\theta) = \sum_{j=n}^{n+m-1} w(\kappa_j(\theta)Q, \theta_j) I(Q)=โˆ’j=nโˆ‘n+mโˆ’1โ€‹pjโ€‹(ฮธ)=j=nโˆ‘n+mโˆ’1โ€‹w(ฮบjโ€‹(ฮธ)Q,ฮธjโ€‹)

The coordinator determines the optimal acceptance ratios ฮทi\eta_iฮทiโ€‹ for each data provider by maximizing the social surplus:

ฮทโˆ—=argโกmaxโกฮทโˆˆ[0,1]dim(di)ร—n{I(d^โ‹…ฮท)โˆ’โˆ‘i=0nโˆ’1ci(d^iโ‹…ฮทi,ฮณ^i)}\eta^* = \arg\max_{\eta \in [0,1]^{dim(d_i) \times n}} \left\{ I(\hat{d} \cdot \eta) - \sum_{i=0}^{n-1} c_i(\hat{d}_i \cdot \eta_i, \hat{\gamma}_i) \right\}ฮทโˆ—=argฮทโˆˆ[0,1]dim(diโ€‹)ร—nmaxโ€‹{I(d^โ‹…ฮท)โˆ’i=0โˆ‘nโˆ’1โ€‹ciโ€‹(d^iโ€‹โ‹…ฮทiโ€‹,ฮณ^โ€‹iโ€‹)}

Data providers contribute their accepted datasets d^โ‹…ฮทโˆ—\hat{d} \cdot \eta^* d^โ‹…ฮทโˆ— to distributed machine learning. If a data provider cannot contribute d^iโ‹…ฮทiโˆ—โ‰คd^i\hat{d}_i \cdot \eta^*_i \leq \hat{d}_id^iโ€‹โ‹…ฮทiโˆ—โ€‹โ‰คd^iโ€‹, a high punishment is imposed. The income to the federation is I(d^โ‹…ฮทโˆ—)I(\hat{d} \cdot \eta^*)I(d^โ‹…ฮทโˆ—).

The PVCG payment pi(โ‹…)p_i(\cdot) piโ€‹(โ‹…) consists of the VCG payment ฯ„i\tau_iฯ„iโ€‹ and the optimal adjustment payment hiโˆ—h^*_i hiโˆ—โ€‹:

pi(โ‹…)=ฯ„i(โ‹…)+hiโˆ—(โ‹…)p_i(\cdot) = \tau_i(\cdot) + h^*_i(\cdot)piโ€‹(โ‹…)=ฯ„iโ€‹(โ‹…)+hiโˆ—โ€‹(โ‹…)

The VCG payment ฯ„i\tau_i ฯ„iโ€‹ to data provider iii is calculated as:

ฯ„i=S(d^,ฮณ^)โˆ’Sโˆ’i(d^โˆ’i,ฮณ^โˆ’i)+c(d^iโ‹…ฮทiโˆ—) \tau_i = S^(\hat{d}, \hat{\gamma}) - S^{-i}(\hat{d}{-i}, \hat{\gamma}_{-i}) + c(\hat{d}_i \cdot \eta^*_i) ฯ„iโ€‹=S(d^,ฮณ^โ€‹)โˆ’Sโˆ’i(d^โˆ’i,ฮณ^โ€‹โˆ’iโ€‹)+c(d^iโ€‹โ‹…ฮทiโˆ—โ€‹)

where Sโˆ—S^* Sโˆ— represents the maximum producer surplus and Sโˆ—โˆ’iS^*{-i} Sโˆ—โˆ’i is the surplus without data provider iii. d^โˆ’i\hat{d}{-i}d^โˆ’i and ฮณ^โˆ’i\hat{\gamma}_{-i}ฮณ^โ€‹โˆ’iโ€‹ denote the claimed datasets and reported cost types excluding data provider iii.

๐ŸŽฎ