Centralized or Decentralized? Value Proposition of DML — On-Device Machine Learning

Published in

Decentralized Machine Learning

4 min readApr 8, 2018

Value Proposition of DML — On-Device Machine Learning

We have discussed how DML Infrastructure may revolutionize the current big data and machine learning market. In this article, we will further highlight the major differences between traditional centralized solutions of machine learning and the decentralized features of DML Infrastructure design. Besides, we will discuss the potential cooperation opportunities among different projects and parties across the decentralized space, and how different contributors of machine learning development may be complemented under a decentralized and democratic environment.

The table below summarizes the differences between centralized and decentralized solutions for machine learning development from various key perspective such as data, processing power, algorithm development and algorithm training.

Comparison of Centralized and Decentralized Machine Learning

Value Proposition of DML — An Infrastructure to Utilize Decentralized Data, Processing Power and Algorithms for Machine Learning Development

DML aims to create an infrastructure for decentralized machine learning development by unleashing the potential of untapped data with data privacy protected, leveraging unutilized processing power and encouraging algorithm development through innovation from periphery. In light of this, DML protocol is intended to be comprised of various layers of features:

DML Mobile App — a mobile application to facilitate on-device machine learning to be run on individual devices;
Decentralized Nodes — nodes to facilitate algorithms delivery, results aggregating and averaging through federating learning, prediction reports generation and algorithm fine-tuning;
DML Algo Marketplace — an algorithm marketplace for algorithm developers to create and market their algorithms and to crowd-source collective resources for algorithm fine-tuning

With the aid of DML protocol, an ecosystem that connects potentially billions of devices, tens of thousands of machine learning developers and millions of customers who wishes to apply machine learning is aimed to establish.

The potential usage of DML Infrastructure is numerous. This may include market research, resources planning, investment sentiment analysis, consumer behaviour analysis and political assessment etc. Parties such as companies, research institutions, government, non-government organizations or even individuals will be able to acquire machine learning services through the marketplace. Besides, anyone who has a smart device can participate and be rewarded by contributing idle processing power and data usage for machine learning development with privacy protected.

Cooperation Opportunities with Other Decentralized Solutions

Data Marketplaces

In a decentralized data marketplace, data will be encrypted, uploaded and stored in some distributed file storage nodes. Such dataset can be used as a common dataset for the crowd-sourced machine learning trainers for fine-tuning the machine learning algorithms. Besides, such dataset can facilitate to organize algorithm competitions to locate talents and source valuable algorithms.

Computation Power Networks

Some computer owners are trading their idle processing power through computation power networks. This powerful resource can also fuel the development of DML Infrastructure. While on-device machine learning will distribute part of machine learning processing on device, the decentralized nodes could be run in such distributing computing network to avoid single point of failure as well as making use of distributed resources.

Algo Marketplaces

The machine learning algorithms developed in other marketplaces or in house may be deployed in the coming DML Infrastructure to access distributed data on-device for machine learning prediction. DML also aims to support the fine-tuning flow of machine learning algorithms by crowdsourcing machine learning trainers and richer training datasets. Therefore, those algorithms can make use of the coming DML Infrastructure for fine-tuning to make them better and of higher accuracy and thus increasing their value and usefulness.

Return Data Autonomy to the People

With the view to protect data privacy, DML Protocol will be designed not to extract raw data from the devices of data owners. Data owners will be able to choose specific type of data that are allowed to be run by machine learning algorithms. Besides, only machine learning results will be delivered after aggregation and averaging by the nodes with the aid of federated learning.

Wrapping Up

Tech giants are gaining control over majority of individual private data without providing a fair compensation. Besides, individual may lose data autonomy if one is relying on centralized storage and solution. DML’s vision is to allow data owners to gain back data ownership and control over their own data while be able to contribute to and benefit from machine learning development. To understand more about DML Infrastructure, please refer to our whitepaper, video and other materials.