Welcome to the 0G Research
At 0G, we bridge the gap between advanced AI and Web3 technologies.
Our research department is dedicated to pioneering breakthroughs in
decentralized AI, fostering collaboration, and driving innovation
across blockchain ecosystems. Below are the major research areas we
are currently pursuing.
Optimization of Model Training in Decentralized AI Systems
In this task, we aim to design a suite of methodologies and frameworks
at both the system and algorithm levels, to boost training efficiency
in large-scale decentralized environments. Specifically, we will
consider the following optimization directions:
Communication Optimization
Decentralized learning requires participating nodes to frequently
share intermediate training results via networks. Larger models and
data could result in huge communication costs, affecting overall
training efficiency. We will introduce cutting-edge solutions to
reduce such overhead, including lossless gradient compression and
quantization.
Local Computation Optimization
Training models at scale incurs significant computation overhead and
resource consumption. This becomes a bottleneck for local devices in
decentralized AI. We aim to alleviate such pressure from two
perspectives: integrating advanced training pipeline and
parallelization technologies to accelerate the training process, and
pruning the models to reduce computation, thus realizing efficient
training at the edge.
Performance Optimization with Heterogeneous Data
Data heterogeneity across different nodes in decentralized systems can
deteriorate model accuracy due to local model drift. To fix this issue
in practical data-heterogeneity scenarios, we will design novel
algorithms for training and aggregation. For instance, we can
adaptively adjust the learning rate of each neuron based on local data
to rectify model drift, and perform meticulous round selection for
more robust aggregation.
Performance Optimization with Dynamic Environment
In a practical decentralized AI system, nodes can join or exit the
collaborative tasks in an unpredictable way. Robust solutions are
required to handle such dynamic features while guaranteeing model
performance. To this end, we will design new asynchronous learning
protocols and schedulers to efficiently and accurately predict node
participations, and select high-quality updates from certain nodes and
iterations for enhanced aggregation.
Model Alignment in Decentralized AI Systems
Contemporary large language models demonstrate exceptional text
interpretation and generation capabilities. However, they also raise
ethical risks as they could inadvertently generate inappropriate,
biased, harmful, or non-factual content. These risks are exacerbated
in decentralized AI ecosystems, where each node’s training data is
neither controllable nor filtered. It is important to perform model
alignment, ensuring the model output aligns with human values. We will
make several efforts toward this task.
Enhanced Learning Algorithms with Human Preference
The most common solution for alignment is to integrate human
preferences as human values in model optimization, e.g., Reinforcement
Learning from Human Feedback (RLHF) and Direct Preference Optimization
(DPO). We aim to apply these strategies to decentralized settings and
make them more efficient. Possible improvements include the
construction of higher-quality preference datasets, design of advanced
reward functions, and distributed sharing of human feedback across
different participating nodes.
Self-Regulation and Correction
In social psychology, perspective taking is an important emotional
intelligence skill, inspiring individuals to leverage self-awareness
for behavior regulation. Inspired by this principle, we will propose
new alignment strategies, which guide the model to automatically
inspect its output responses, identify any content misaligned with
human values, and rectify it. A new end-to-end regulation and
correction pipeline with advanced prompts will be established to
achieve this goal.
Decentralized Debating for Alignment
Debating is another popular alignment method, where multiple models
(or agents) debate with each other to produce the most accurate and
valuable content. This approach is a natural fit for decentralized AI,
where there are multiple models from different nodes ready for debate.
We will implement this strategy in real-world, large-scale
decentralized scenarios and make adaptations to enhance alignment
efficiency and effectiveness.
New Blockchain System Empowered by Multi-Agent Technology
LLM-based multi-agent systems are becoming prevalent due to their
excellent capability of handling complex tasks. Such systems
coordinate distributed agents—each with specialized AI models—to
collaborate on any given task. The characteristics of these systems
align well with blockchain environments, indicating a significant
potential for integration to enhance both functionality and
efficiency. Specifically, each node in a blockchain can be implemented
as an agent to establish a multi-agent setup. Following this
principle, we will consider several applications for multi-agent-based
blockchain:
Smart Contract Management
Smart contracts are a critical component in blockchain to automate
transaction execution. It is promising to analyze, manage, and
optimize this software in a distributed manner. We could implement a
multi-agent solution in the blockchain, with each agent focusing on
different functionalities. Their collaboration could significantly
augment the blockchain with comprehensive smart contract services.
Anomaly Detection
During blockchain execution, malicious entities may attempt to
interfere with transactions, consensus mechanisms, or cross-node
communications. It is thus vital to introduce security schemes to
monitor the system and detect any anomalies. We will design and
develop multi-agent systems to achieve this goal. By encouraging
different agents to focus on various aspects of events and
coordinating their decisions, the trustworthiness of the blockchain
environment will be greatly enhanced.
Recent Publications
Below is one of our latest publications that showcases our ongoing
research:
BadSFL: Backdoor Attack against Scaffold Federated Learning
Authors: Xingshuo Han, Xuanye Zhang, Xiang Lan,
Haozhao Wang, Shengmin Xu, Shen Ren, Jason Zeng, Ming Wu, Michael
Heinrich, Tianwei Zhang
Year: 2024
arXiv ePrint: 2411.16167
URL:
https://arxiv.org/abs/2411.16167
BibTex Citation
@misc{han2024badsflbackdoorattackscaffold,
title={BadSFL: Backdoor Attack against Scaffold Federated Learning},
author={Xingshuo Han and Xuanye Zhang and Xiang Lan and Haozhao Wang and Shengmin Xu and Shen Ren and Jason Zeng and Ming Wu and Michael Heinrich and Tianwei Zhang},
year={2024},
eprint={2411.16167},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2411.16167},
}