TensorFlow & PyTorch User Group Talks [Two 20-minute talks]
Thursday, October 17, 4:30–5:30 pm, 138 Lewis Science Library
[Pizza will be provided. RSVP now!]
- Accelerating automated modeling and design with stochastic optimization and neural networks (20-minute talk)
Alex Beatson, Graduate Student, Computer Science
Faculty Advisor: Ryan Adams, Princeton University
For tasks such as learning to learn, identifying the parameters of natural systems, or optimizing the design of mechanical parts, tuning the (hyper)parameters of a system can require running a high fidelity numerical method at each optimization step. I will discuss two methods which aim to accelerate automated modelling and design by reducing this computational cost. The first is “Randomized Telescope” gradient estimators, which provide cheap unbiased stochastic gradients for problems where the objective is the limit of a sequence of increasingly costly approximations. These can accelerate tasks such as optimizing hyperparameters of neural networks and fitting parameters of ODEs. The second is “Neural Model Order Reduction”, which uses deep learning and integrates PyTorch and Fenics (an open source PDE solver) to reduce the dimension of nonsmooth PDEs. Our preliminary work uses this to efficiently simulate mechanical metamaterials: materials engineered with fine-scale structure which can be expensive to simulate but which gives rise to macroscopic properties not found in nature.
- Modeling Human Sequential Decision-Making (20-minute talk)
Mark Ho, Postdoctoral Researcher, Computational Cognitive Science Lab, Department of Psychology
Planning complex sequences of decisions in dynamic environments is challenging, but people are quite good at this. For instance, completing a Ph.D. requires planning at multiple levels of abstraction, from deciding a dissertation topic to attending classes to figuring out what to eat for breakfast. Thus, a major challenge for psychology and neuroscience is identifying, in computational terms, what mental strategies enable people to construct and implement long term plans. In this talk, I will discuss how we have used PyTorch to model the dynamics that emerge from a person interacting with a task who is both planning and "meta-planning" in order to accomplish their goals.
A lightning talk will also be added. If you would like to give a presentation, please send an email to firstname.lastname@example.org.
The user group is sponsored by the Princeton Institute for Computational Science and Engineering (PICSciE) and the Center for Statistics and Machine Learning (CSML).
Questions? Email email@example.com