A Peek into the Learning Algorithms Behind Artificial Intelligence
In my research, I aim to unify fundamental tools from systems theory, control, signal processing, and optimization to better understand the behavior of key algorithms used in artificial intelligence, particularly those used to train deep neural networks.
More broadly, my goal is to demystify the black-box, heuristic nature of modern learning algorithms by developing principled interpretations that improve transparency, stability, and interpretability.
Doctoral Research (Artificial Intelligence)
My work focuses on stochastic gradient-based learning for training deep neural networks.
A central question is:
How can we move from heuristic algorithm design to a principled, unified algrorithimic framework that automatically adapts and optimizes learning dynamics?
Momentum as Lowpass Filter Design
Trust-Region Optimal Learning Rate Principle
Annealing as a Variational Trust-Region Window Problem
A Memoryless Proxy for Lowpass Regularization
An AutoSGM Framework: Conjugate Gradient Method
Dissertation: Fundamental Signal Processing Elements in Accelerated Learning
Master’s Research (Control Theory)
This work focused on the dilemma of tuning proportional–integral–derivative (PID) controllers.
Future direction:
How can we develop a unified framework that bridges optimization and control theory for algorithm design?
Dillema of PID Tuning PID Project View on GitHub