Background
Assisted learning is a machine learning paradigm that allows decentralized learning agents to autonomously improve each other’s learning capability without transmitting local data, models, and even learning objectives.
This framework differs from classical distributed machine learning and the recent Federated learning, which often assume a global training model or objective known to all the learners. Assisted learning is particularly suitable for collaborative learning across organizational learners with rich resources, e.g., research teams, Internet of Things, AI business units, and government agencies.
Concept
Assisted learning is a machine learning paradigm that is characterized by:
- Autonomy: There is no central controller that orchestrates the learning and optimization
- Model privacy: Neither the agent being assisted nor the assisting agents will share their private local models
- Data privacy: Neither the agent being assisted nor the assisting agents will share their private data
- Unlimited local resources but limited assistance: The agent being assisted has a limited budget to purchase assistance services from other learners. But each agent has rich resources (e.g., data, model, and computation) in its locality.
The goal of the Assisted learning protocol is to significantly expand the single-agent learning capabilities with non-private information exchanges. The above features distinguish Assisted learning from the related topic of Federated Learning, a distributed learning framework that features communication efficiency. The main idea of Federated Learning to learn a joint model using the averaging of locally learned model parameters so that the training data do not need to be transmitted. Conceptually, the main objective of Federated learning is to exploit computation and data resources of edge devices to achieve a global objective, while the goal of Assisted learning is to provide protocols for decentralized agents to assist each other with their private modeling processes. Methodologically, in Federated learning, a central controller orchestrates the learning and the optimization, and Assisted Learning provides a protocol for the agents to optimize and learn among themselves autonomously.
Application
Assisted learning is particularly suitable for multi-organizational, autonomous, and decentralized learning scenarios, where each agent is often equipped with rich computation resources, unique data, and a sophisticated model. An agent, say A, lacks some side information and needs other agents to “assist” it. However, none of the agents will disclose their private data, model, or learning objective to others.
A medical institute may be helped by multiple clinical laboratories and pharmaceutical entities to improve clinical treatment and facilitate scientific research. Financial organizations may collaborate with universities and insurance companies to predict loan default rates. The organizations can match the correspondence with common identifiers, such as user identification associated with the registration of different online platforms, timestamps associated with different clinics and health providers, and geo-locations associated with map-related traffic and agricultural data. With the help of our framework, they can form a community of shared interest to provide better Machine-Learning-as-a-Service (MLaaS) without transmitting their local data, models, and objective functions.
Goal
To significantly enhance single-organization learning capabilities through private assistance from other heterogeneous agents:
Organizations with rich resources
Smart IoT devices with different capabilities
Research teams in various fields
Companies targeting similar customers
Challenge
Agents are privacy-sensitive.
They will not transmit local proprietary:
-
Data
-
Model
-
Learning procedure
-
Learning objective
Decentralized Multi-Target Cross-Domain Recommendation for Multi-Organization Collaborations
We propose Decentralized Multi-Target Cross-Domain Recommendation (DMTCDR) with Multi-Target Assisted...
Read MoreGAL: Gradient Assisted Learning for Decentralized Multi-Organization Collaborations
In this framework, all participants collaboratively optimize the aggregate of...
Read MoreParallel Assisted Learning
We develop a general method called Parallel Assisted Learning that...
Read MoreAssisted Learning for Organizations with Limited Imbalanced Data
In our assisted learning framework, an organizational learner purchases assistance...
Read MoreGet in Touch
Communicate Colloborate Celebrate
Phone
(+1) 617-999-3243
dingj@umn.edu
Address
224 Church Street, Minneapolis