By Hung Viet Tran
We are living in a digital ocean. From the complex patterns of our climate and the subtle signals in our health data to the vast networks of the global economy, we are surrounded by a tidal wave of information. This “big data” holds the keys to solving some of humanity’s greatest challenges. But how do we unlock its secrets?
Imagine trying to read every book in the world’s largest library before you could answer a single question. That’s the challenge traditional data analysis faces. It is slow, cumbersome, and often, the answers it provides are confusing. An AI might tell us what to look at, but it rarely explains why. This “black box” approach is a major bottleneck, especially when discovery is a trial-and-test process.
Our PhD project is dedicated to changing that. We are building the next generation of artificial intelligence tools that are not only powerful but also interactive, transparent, and computationally efficient. Our mission is to transform AI from a mysterious black box into a collaborative partner that can help us navigate the data ocean.
We focus on two core pillars to make this happen:
- Making AI Interactive and Intelligent: Traditional algorithms often require all data to be available before they can even start, which is a huge problem for massive, ever-growing datasets. We’re building algorithms that can learn incrementally, processing data as it arrives. By combining this with the power of parallel computing, we teach AI to use all of its brainpower at once, making the analysis fast enough for a human to interact with, test ideas, and get results in near real-time.
- Making AI Clear and Explainable: An answer without an explanation is only half a solution. If an AI flags a patient as high-risk or identifies a financial anomaly, experts need to know why. Our work focuses on building explainability directly into our algorithms, so they can provide detailed reasons for their conclusions, making them far more useful and trustworthy in practical, real-world scenarios.
Our Progress in Action
This is not just a theoretical goal; we’re already making exciting progress. Our recent work, highlighted in three research papers, shows how we’re turning this vision into reality.
1. A Leap in Forecasting Power: WaveletMixer
Predicting long-term trends is a classic “big data” challenge where traditional methods struggle. Our paper, “WaveletMixer: A Multi-Resolution Wavelets Based MLP-Mixer for Multivariate Long-Term Time Series Forecasting, AAAI, 2025” introduces a novel and highly intelligent algorithm for this task. By analyzing data at multiple resolutions simultaneously, like a camera that can see both the forest and the trees at the same time, WaveletMixer achieves outstanding performance in forecasting. This work is a prime example of our goal to develop new, powerful data mining techniques capable of tackling complex, large-scale problems.
2. Giving AI a Voice to Explain Itself: InteDisUX
Directly addressing our second pillar, the paper “InteDisUX: Interpretation-Guided Discriminative User-Centric Explanation for Time Series, AAAI, 2025” is all about breaking open the black box. When an AI model classifies a piece of time-series data, like an ECG heartbeat signal, InteDisUX can pinpoint the exact segment of the signal that lead to the decision.This provides users with clear, meaningful insights, which is crucial for building trust and enabling practical use in sensitive fields like healthcare.
3. Efficient and Practical by Design: Random Erasing vs. Model Inversion
Part of building intelligent systems is ensuring they are robust, practical, and computationally efficient. Our third paper, “Random Erasing vs. Model Inversion: A Promising Defense or a False Hope?, Just Accepted at Transactions on Machine Learning Research” demonstrates this principle. While exploring the critical issue of data privacy, we did not just find a solution; we found an incredibly efficient one. The paper shows that a simple, computationally cheap technique of “random erasing” parts of data during training can serve as a powerful defense against privacy attack. This work highlights our project’s core focus on developing novel solutions that are not just effective, but also simple and fast enough for practical, large-scale deployment.
The Road Ahead
From helping scientists model climate change to giving doctors tools to better diagnose diseases like Obstructive Sleep Apnea, the applications of interactive, explainable, and efficient AI are limitless. These papers are just the beginning of our journey. We are committed to continuing this work, pushing the boundaries of what’s possible, and ensuring that as our world becomes more data-driven, we are building a future where technology truly serves humanity.