Deep Learning for High-Quality Spatiotemporal Prediction

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Deep Learning for High-Quality Spatiotemporal Prediction"

By

Mr. Jiachuan WANG


Abstract:

Spatiotemporal prediction has garnered significant attention for many years. In
recent years, deep learning methods have emerged as effective models for
spatiotemporal data, surpassing traditional methods in tasks such as data
enhancement and prediction. While considerable effort has been dedicated to
developing deep learning methods for spatiotemporal data mining, its popularity
has soared due to the increasing need for high accuracy in managing vast
amounts of data in the big data era. Besides, complex application scenarios
call for robust method designs that balance professionalism and versatility.

This thesis aims to propose a pipeline for high-quality spatiotemporal
prediction. The pipeline includes data enhancement as preprocessing and the
design of effective tools for data modeling and prediction. In the first
problem, our aim is to denoise noisy input frames such as radar images in
situations where clean reference samples are not available, which is a common
issue in spatiotemporal data collection. Constructing it as an image denoising
problem, we use a self-supervised deep learning method that optimizes models
based solely on noisy images. The performance of this method heavily depends on
the scale of the noise, which is often unknown. To address this issue, we
closely estimate an upper bound for the noise scale without relying on clean
images. Additionally, we design a framework for iterative model updates and
scale estimation. For the second problem, we first introduce meta operators to
capture complex and abstract high-order dynamics. This is particularly
important for spatiotemporal data, such as rainfall systems, compared to simple
image processing. Existing methods utilize activation functions to model
nonlinearity, but these functions have bounded first-order derivatives. Taking
inspiration from classical dynamic system simulation, we apply polynomial
activation functions as a powerful tool for modeling high-order non-linearity.
To further improve training robustness, we propose a Range Norm and apply it
before the activation functions. Next, we develop an effective module to
uncover underlying principles within complex correlations. Since large models
designed for spatiotemporal prediction tend to learn inherent principles of
dynamic systems slowly and expensively, we explore the variation of patterns
across different scales. We hypothesize that the decrease in monosemanticity
could be the reason for the model's good performance. To address this, we
propose a framework that actively detects monosemantic neurons and inhibits
them to induce the emergent abilities of models. With our solution, inputs are
decomposed, and outputs are inferred based on complex and abstract hidden
patterns. Finally, extensive experiments on real and synthetic datasets
demonstrate the effectiveness of our proposed solutions.


Date:                   Monday, 15 January 2024

Time:                   4:30pm - 6:30pm

Venue:                  Room 3494
                        Lifts 25/26

Chairman:               Prof. King Lun YEUNG (CBE)

Committee Members:      Prof. Lei CHEN (Supervisor)
                        Prof. Charles NG (Supervisor, CIVL)
                        Prof. Qiong LUO
                        Prof. Dan XU
                        Prof. Can YANG (MATH)
                        Prof. Haibo HU (HKPU)


**** ALL are Welcome ****