Projects
1. Distributed Learning
Distributed learning has emerged as a transformative technological trend, gaining remarkable momentum over the past decade. This paradigm shift in artificial intelligence, networking, communication, information theory, and machine learning applications has far-reaching implications, spanning a diverse spectrum of fields. It encompasses the efficient training of large-scale AI models, exemplified by the success of ChatGPT, as well as collaborative learning strategies crucial for the development of future autonomous driving systems. Moreover, its impact extends to precision healthcare and the integration of federated wearable devices, revolutionizing how we approach these domains. Within this dynamic landscape, several critical factors come into play, shaping the effectiveness and efficiency of distributed learning systems: 1) resource constraints, such as limitations in communication, computation, storage, and caching capabilities; 2) the inherent resource and data heterogeneity of distributed systems; 3) the imperative of preserving privacy; 4) ensuring robustness; and 5) context awareness, an essential element for adapting to changing environments and scenarios. The intricate interplay of these components directly influences the performance of distributed learning systems.
Our research endeavors are dedicated to dissecting and understanding these intricate interplays. We aim to construct a robust theoretical framework that provides fundamental insights into the dynamics of distributed learning systems. Our approach involves the analysis of fundamental tradeoffs among different factors, e.g., interactions between communication and computation, derivation of performance bounds, the development of tailored algorithms, and continuous refinement of our theoretical framework based on empirical insights and the deployment of proposed algorithms.
Specifically, our research project encompasses
1.1 Communication-efficient learning and fundamental tradeoffs between communication and computation. This includes the design of efficient coding frameworks, e.g., pliable index coding, and the compression of AI model parameters for efficient communication.
Coding for wireless content transmission
Coding for data shuffling
Adaptive parameter compression (communication-efficiency) for distributed learning
Matrix factorization-based communication-efficient distributed learning
1.2 Robust distributed learning. This includes coded computation and statistical methods for mitigating stragglers and combating adversary attacks.
Adversary attacks
Data encoding for mitigating adversary attacks
Live gradient compensation for evading stragglers
Coded alternating least square for LLM trainining
2. Recommender systems and reinforcement learning.
Recommender systems will become the future information acquisition technology by allowing users to obtain content based on their intent other than explicitly searching for prespecified content. In the first part of this project, we are developing learning algorithms to realize effective and accurate recommendations. We are conducting experiments over several real data set, such as Yahoo! news and ads recommendations. In the second part of this project, we are trying to build up a bandwidth-aware recommender system that can adjust the recommendation strategy based on current bandwidth of the system. We try to build fundamental framework for the learning and analysis of the system.
Some interesting topics I am currently working on include reinforcement learning, contextual learning, combination of expert advices, and wireless recommender systems.
3. Distributed machine learning and groups of drones.
In order to handle large volumes of data, distributed computing is widely used for machine learning and optimization tasks. Communications among the distributed nodes sometimes become the key bottleneck in terms of runtime performance, especially for the edge computing system where wireless bandwidth resource is scarce. We try to leverage the redundant resources that are available to the system (such as storage resources, computing resources, and power resources) to tradeoff the communication bottleneck. We try to design different approaches and analyze the performance of the system when conducting different tasks.
One particular application area is to study the group of drones that involve the interactions between communication and learning.
4. Novel communication paradigm for big data traffic.
With the booming of big data applications, the processing of big data becomes mature. However, there is little known about the communications of big data. Today's communication networks increasingly deliver big data content rather than traditional specific messages. Popular networks that serve content-type traffic include advertising networks, news aggregators, search engines, recommender systems, and social media. This research direction formulates a novel theoretical framework aiming to determine the fundamental performance limits and design principles for big data generated content-type networks. This new research promotes the progress of science and has the potential to transform the way big data generated content-type traffic is encoded and transmitted in networks. As a result, this research is expected to benefit society at large by laying the foundation for a more efficient design of network services, and to be of immediate and far-reaching use for both private and public sectors.
5. Natural language processing.
This project is about event-based natural language understanding and the applications on FinTech, reasoning, and other subsystems.
6. FinTech.
This project is about using various machine learning methods for financial technology, such as report analysis, risk prediction, and stock price prediction.
7. Other machine learning theory and applications.
This project is about analyzing CNN, GAN, DQN, etc. and applications (such as ad placement, smart grids, sharing economy) with them.