Research Interests in detail

Research Vision

I am intended to develop and build intelligent agents that can learn knowledge based on their interaction with the environment and finish long-term complex tasks without human intervention. And I believe reproducing human intelligence step by step and building intelligent robots will be a reality in the foreseeable future.

On the other hand, I should ask "What intelligence does this project is trying to reveal?" in all my research projects.

What social impact my research vision could have?

Recently, DeepMind worked on generating data that helps mathematicians gain intuition through machine learning. The "generating intuition" process is still done by humans.

I believe my work will start the process of generating intuition using artificial intelligence, which takes the place of scientists in specific areas.

My attitude towards challenging math theory in Machine Learning

Considering I took Matrix Analysis (graduate level) and Convex Optimization (graduate level) and I got As from them, I have the confidence to learn and work on the math problems in the ML theory. Recent research experiences did put me a lot on engineering work instead of math theory.

But I am also happy to gain insights from math about how we can get closer to human intelligence.

My attitude towards Deep Learning in Computer Vision research

I don't like to do only Computer Vision or Graphics research. My research vision is mainly about intelligence. I understand that AGI like the Godel machine is hard to implement.

Thus, I am willing to do a Computer Vision project only if this project is a preliminary part of a big artificial intelligence project or this project is a necessary practice for me to gain deep learning experience or mathematical skills.

Why I don't study psychology or neural science?

I just do not major in psychology or neural science. I have taken "Fundamental Psychology Theories and Methodology" as in my undergraduate transcript. I have also gone to several neural science seminars held by ShanghaiTech, which gave me the impression of how neural science has developed in terms of human intelligence.

Furthermore, my engineering background helps me accomplish more in Computer Science, and the research progress in intelligence behavior from either psychology or neural science still requires coding to reproduce them.

General Research Framework (a.k.a SciBot project)

(You may consider it as a research proposal draft)

1. To achieve Artificial General Intelligence, working on robot learning is the most practical path.

2. To work on robot learning as a path toward AGI, two major aspects need to be studied:

   1. Unsupervised method to learn general world model and causal relations between all entities in the world. (Reasoning part)

   2. General control algorithms that let the robot act in the real world in order to do experiments on its own. (Action part)

Reasoning part

1. An unsupervised method that extracts abstract entities from the world. For example, defining subtasks based on the long-term goal, emerging ego-centric representation for the robot, or extracting "sub-skills" from the expert demonstrations.

Thus, I am currently studying and working on unsupervised object-centric representation learning through videos.

2. An unsupervised method to generate a relation graph between the extracted entities. The system should generate latent variables when needed, with minimum human supervision.

3. An unsupervised learning algorithm that learns to generate dynamic causal graphs based on its past experiences with the objects in the world. This involves experiments designed by the embodied agent and counter-factual reasoning and assumptions.

Action part

1. An embodied agent that can move like a human requires general control algorithms for the robot. For example, a control algorithm for a legged robot that can go through almost any terrain.

CPG could be another solution to RL methods

2. A general dynamical model that helps the robot to manipulate objects. For example, picking any objects within its hardware ability.