Homepage of Dr. Hung Le

About

I am Hưng, a research lecturer and Australian Research Council DECRA Fellow at Applied Artificial Intelligence Institute (A2I2), Deakin University.

News

My Research

1. Deep Neural Network Models with Artificial Neural Memory:

• Created novel architectures that incorporate artificial neural memory, leading to breakthroughs in multi-modal and generative AI. ✏️ KDD18, NeurIPS18, ICLR22
• Developed foundational theories for memory operations, enabling more effective learning and reasoning in neural networks. ✏️ ICLR19, ICML20, ICLR23

2. General-Purpose and Human-Aligned Neural Architectures:

• Investigated the design and implementation of neural computers akin to Universal Turing Machines, capable of handling diverse tasks such as continual learning, machine reasoning and reinforcement learning. ✏️ ICLR20, ICML22
• Devised a general memory module that enhances the generalization and reasoning capabilities of various deep learning models from small RNNs to giant LLMs. ✏️ TMLR24
• Developed alignment strategies for LLMs, leveraging collective models or program memories to enhance model generalization and safety. ✏️ AAAI25

3. Memory-Based Reinforcement Learning Agents:

• Designed agents that leverage memory to improve decision-making. ✏️ NeurIPS21, AAMAS24
• Developed memory-based optimization approaches to improve reinforcement learning. ✏️ NeurIPS22, AAAI22

1. Complex Time-series System Forecasting Reinforced by Expert Knowledge

This project leverages LLMs and memory techniques with reinforcement learning to integrate expert knowledge for time-series forecasting. It aims to enhance predictive accuracy in critical areas like healthcare and material science, driving informed decision-making and fostering advancements in forecasting models. This project has been awarded a DECRA25 grant of $428,331 AUD.

2. Obedient Large Language Models (LLMs)

As LLMs rapidly advances, the need for obedient LLMs is urgent—without it, these models can generate outputs that ignore user needs, values, or ethics. LLMs must be designed to listen and follow orders precisely. By incorporating techniques like uncertainty estimation and alignment strategies, LLMs can be trained to reliably follow human instructions, ensuring their outputs align with societal expectations and ethical standards, while reducing risks and ensuring more controlled, effective human-ai symbiosis. This project has secured a valuable seed investment of $10,000 USD through Cohere For AI Research Grant.

Past Research Projects

• Memory and Attention in Deep Learning (my PhD at Deakin University, 2017-2020) PhD. Thesis. My great supervisors: A/Prof. Truyen Tran and Prof. Svetha Venkatesh
• Graphical & Topic Model (when I did Master at HUST with A/Prof. Khoat Than, 2016-2017 dropout)
• Kalman Filter & Optimal Control (when I worked at Viettel R&D, 2015-2017)
• Fuzzy Logics (when I did Bachelor Honours at HUST with A/Prof. Khang Tran, 2015)