Leaving the friendly confines of the Redwood Center for a truly strange wider world this semester are three PhD students we will all dearly miss.
First up we have Michael Fang, esteemed denizen of Evans 563 and scientific polyglot, whose thesis addressed the problem of using analog computers for machine learning—specifically strategies for managing analog noise. One of his projects, a collaboration with Intel, focused on mitigating the effects of fabrication uncertainty on optical neural networks. Another project, taking a constructive rather than a reductive approach, designed a system in which inherent device stochasticity was used for sampling in order to train latent variable models. Michael was advised by Mike DeWeese and Bruno Olshausen. He has founded and is currently running a highly mysterious stealth start-up company.
Charles Frye, a longtime spark of color and pedagogical clarity around Redwood, developed thesis work focusing on the geometric structure of the loss functions of neural networks. In particular, his work, with fellow graduate students Neha Wadia, Jamie Simon, and Andrew Ligeralde, demonstrated that a large class of methods used to understand the curvature of neural network loss functions suffered from a previously unnoticed flaw. These findings call into question the prevailing conventional wisdom that neural networks are easily trained due to the absence of bad local minima. Charles was advised by Mike DeWeese and Kris Bouchard. He will be joining Weights & Biases as a Deep Learning Instructor—Weights & Biases is a San Francisco-based start-up that produces tools for collaborative, reproducible machine learning research and industrial application. Charles will be creating in-person and online educational content (e.g. YouTube lectures).
Last but certainly not least we have Ryan Zarcone, whose laughter is probably still ringing in Evans Hall (if only we were allowed to go inside). Ryan’s thesis work examined the problem of storing analog information in a set of stochastic analog devices. Specifically, his work with Redwood members Alex Anderson, Dylan Paiton, and Jesse Engel cast the problem of data storage as a communication problem and then used tools from information theory and machine learning to construct reliable, learned methods of storing analog data, such as images, in arrays of these noisy devices. This work was a joint collaboration between the Redwood Center and members of H.S. Philip Wong’s Nanoelectronics Lab at Stanford. Ryan was advised by Bruno Olshausen and Mike DeWeese. Ryan will be joining Fountain Therapeutics as a Data Scientist. Fountain Therapeutics is a San Francisco-based biotechnology start-up focused on understanding cellular aging and developing therapeutics to reverse it.
Best wishes from all Redwoodians as these three start their next chapters!