Joey Turnbull
Hello! I'm a physics graduate student at UC Berkeley who studies how neural
networks train. I'm a member of FLAB and work under the supervision of Michael DeWeese and Jamie Simon . I'm currently
working towards a general theory of how ICL and MLP learning dynamics, and more broadly am interested in the properties of generalization in the big data age. I did my undergrad at the University of Washington in physics, where I worked in a number of experimental subfields
(biophysics, gravitation, high-energy, and condensed matter).
In my free time, I am an amateur weightlifter/runner, a sci-fi enthusiast, super-par disc golfer,
a poet, big-time tea lover, and enjoy frolicking in the rain! I also love frogs.
I can be reached at joeyturnbull[at]berkeley[dot]edu! Feel free to say hi!
Research Projects
- Rundown of "Predicting Kernel Regression Learning Curves from only Raw Data Statistics" - the HEA
- Direct measure of DNA bending by quantum magnetic imaging of a nano-mechanical torque-balance arXiv
Quantum mechanics can be used to study microscopic forces!
Blog
Blog - Research
- Linear attention is unable to perform the In-Context Markov Model/In-Context Learning of Representations task
Linear attention is not the minimal model to study ICLR
- Horizontal flips as independent samples - kernels viewpoint of data augmentation
Kernels improve only with *some* data augmentation schemes
- In-browser MLP training
Widget for training MLPs
- Modelscape: Machine learning mega training experiment repo released!
I made a repo that makes training *any* ML model simple!
- MLP experiment repo released!
I made a repo that makes training MLPs quite simple!
- Feature learning networks are/become balanced
Balancedness might be a way to unify all feature learning networks!