Albert Ge

(old)

Hello! My name is Albert Ge. I am currently a master’s candidate at SEAS in Harvard University. I spend time as a graduate researcher at DASlab, advised by Prof. Stratos Idreos, where I think about how to design neural network architectures. By design, we want to determine if there exists a best formulation of parameters to solve a given problem subject to budget, speed, and accuracy constraints. Understanding the space of such architectures will give us better understanding of picking which model to solve the emerging class of cognitive tasks in machine learning, and help us avoid the curse of bigger models = better generalization.

My observation so far is that intuition plays a critical role in this search - we have high-level ideas for how these deep models function, and these ideas scale well to complex tasks. Correct intuition, however, needs to be grounded by theory, and to that end, I also spend time with the Theory of Neural Computation group led by Prof. Cengiz Pehlevan, developing a theory for how we can use less data to achieve better generalization.

Broadly, I’m interested in these questions:

• What is the design space of current neural network models?
• Can we effectively incorporate knowledge from other scientific theories (biologically plausible learning, physical systems) into our artifical models?
• Can these models be scaled on the modern (binary) hardware/software stack?

Before coming to graduate school at Harvard, I had a career as a full-stack software engineer at Academia.edu and at Abbvie Stemcentrx. I obtained my bachelor degree in computer science from Caltech in sunny Pasadena.

I started this website to improve my writing and to keep track of important trends in my research interests. Please feel free to write me at <my-first-and-lastname> at g.harvard.edu if you have any questions or feedback.