10/11/2023
Edgar Duéñez-Guzmán
Title: Studying Social Phenomena with AI
Blurb: In this talk we will talk about how to tackle social challenges using multi-agent reinforcement learning as a scientific tool. We draw insights from game theory and economic models to model relevant phenomena in the real world, using artificial agents as imperfectly-rational decision-makers. We will then talk about how to compile those social study use cases into a benchmark. Melting Pot is an evaluation protocol that compares algorithms that train populations and evaluates the generalization to novel social situations. We are currently running a NeurlIPS Competition, open to all, to improve results on this task suite.
09/13/2023
We will be hosting Dr. Melody Takeuchi, who graduated from Tufts in May 2017, with Christoph Borgers as their advisor. After that, she was an Assistant Professor at Wentworth Institute of Technology before becoming an Internal Consultant at Ab Initio Software.
She will will be briefly talking about her career path. Then she would love to answer any questions about non-research career options. The hope is to have a conversation rather than a proper “presentation.”
In addition, I am one of the main interviewers for my current company, and I’d like to share tips for how to interview well for an industry position.
Her company recruits Tufts STEM PhDs, so if anyone is nearing graduation and interested in having their resume/CV passed along, She is happy to chat about that too.
4/27/2022
Speaker: Dr. David Emerson
Talk Title: A Career Path in Machine Learning with a Graduate Degree in Mathematics
Abstract: In this talk I will discuss my academic background, the research work that I have done as part of the Tufts Scientific Computing group, and my present position as an Applied Scientist at Hyperscience. I will also share some of my experiences with how high-level mathematics is used in the machine learning industry and what skills are most valuable if you’re interested in joining the industry after graduation. In general, I would like the talk to be interactive and will provide significant time for questions about any careers topics of interest.
03/09/2022
Speaker: Benjamin Ellis
Talk Title: Numerical Considerations In a Computational Geometry Engine
Abstract: VulcanForms is a startup developing metal additive manufacturing (3D Printing) technology. Additive manufacturing lies at the crossroads of computational geometry, weldpool physics, and high-precision mechatronics. My primary role has been developing the computational geometry engine — the code that handles all the mathematics needed for other engineers to process CAD models and turn them into physical printed parts. In this talk, I will discuss some of the main considerations needed in order to take abstract mathematical algorithms and implement them in computer architecture, in a way that ensures both high fidelity and high performance.
02/16/2022
Speaker: Dr. Ana Budisa
Talk Title: Computational aspects of modeling brain biomechanics
Abstract: At Simula, we are interested in understanding mechanisms behind biophysical processes in the brain related to aging and neurodegenerative diseases. Recent medical experiments suggest a unified theory on brain fluid dynamics and metabolic waste clearance, but to test these hypotheses we need reliable models and simulations. In this talk, I will present mathematical modeling approaches developed by our research group and my recent work on applicable computational methods and software solutions. The models represent interface-driven multiphysics problems expressed as coupled systems of partial differential equations. However, the complexity of the interface coupling often deteriorates the performance of standard methods to finding the numerical solution. Therefore, we derive scalable solution techniques that properly handle the coupling by using fractional operators. The techniques include, among others, operator preconditioning, multigrid methods, and rational approximation. I will also address how neural networks can be integrated to improve the simulations based on available medical data. Finally, I will demonstrate how those solution methods perform on representative domains, and discuss some advances and challenges when applying to realistic brain geometries.
01/24/2022
Speaker: Sam Polk
Talk Title: MIT Lincoln Lab: History, Present, and Internships
Abstract: The MIT Lincoln Laboratory is a DoD R&D center that has been chartered to develop advanced technologies pertaining to problems within national security. Much of the developed technology has been transferred to industry, academia, and government. This talk will overview a brief overview of the MIT Lincoln Laboratory’s history and origins and then overview some of the important scientific developments that have arisen from Lincoln. I will also overview the Lincoln internship program and talk about how you can apply.
12/1/2021
Speaker: Marshall Mueller
Talk Title: Structured Autoencoders
Abstract: Autoencoders are a form of neural network that tries to both “code” and “decode” a set of data in a way that allows for compressed representations of said data. The coded data points are sometimes referred to as a “latent representation” of the dataset and is often used for downstream tasks such as classification. However, if there is no imposed structure on the autoencoder then there is no guarantee in general that the latent representation will be meaningful, and instead may be an arbitrary memorization of the dataset. To avoid this one must impose some structure to the autoencoder to encourage a desired behavior on the latent space (eg, a linearization of the data set). In the talk we will overview a few of these methods as well as their applications.
11/3/2021
Speaker: Sheng Xu
Talk Title: Machine Learning Against Harmful Ads.
Abstract: Facebook is facing great challenges in its Ads business. To provide a safer environment for user and business clients, auto detection of bad contents is extremely desirable, as the number of Ads has been out of reach by manual reviews. On the other side, a machine learning system, utilizing current development in Transformers and Neural Nets, can detect most of the harmful contents, while a very small amount of uncertainty can be left to human review. We’ll walk through the process and design of such system, and briefly introduce some concepts related. We’ll end with a suggestion on materials that could help you succeed in the interview as well as internship/jobs on the Machine Learning side.
10/13/2021
Speaker: Dr. Peter Ohm from Sandia National Lab
Talk Title: HPC, scientific computing, and multigrid
Abstract: In this talk, I will discuss mathematics and high-performance computing at Sandia National Laboratories. Many applications, such as those arising from geomechanics and aerodynamics, are modeled by increasingly sophisticated and computationally intense numerical models. These numerical models often require the solution to coupled systems of partial differential equations. The development and implementation of efficient and scalable solutions to large-scale multiphysics systems is becoming more and more important as applications become increasingly large and complex. First I will discuss a semi-structured multigrid approach. The semi-structured meshes provide flexibility to address complex computational domains while still allowing most multigrid calculations to be accomplished using efficient structured grid ideas and kernels. Then I will focus on an algebraic monolithic multigrid approach for solving coupled physics systems, with examples from magnetohydrodynamics (MHD).
09/22/2021
Speaker: Matt Friedrichsen
Title: Upgrading TLS for a Post-Quantum World
Abstract: We use multiple forms of encryption to protect our everyday activities online. Most of this is done through a protocol called TLS, that gets used every time you log into a website using https. With commercially available quantum computers coming in potentially a decade (Amazon already provides some quantum computing in the cloud), it’s time to start thinking about how safely encrypt our data against future attacks. We will talk about the different parts of TLS and how it can be modified to use quantum resistant cryptography.