Tom Bertalan
I'm a Data scientist in Amgen's Transformative Digital Capabilities group in Cambridge, MA. I received my PhD from Princeton University's department of Chemical and Biological Engineering, followed by research and engineering work at MIT, JHU, and UMass Lowell.
My research interests are in hybrid modeling, neural system identification, and neurosymbolic computing; with applications in pharmaceutical and chemical processing, and robotics.
Unless otherwise noted, work described here is my own, and does not represent the views of my employer or any other organization.
Visual Scene Grammars
Foundation models can analyze and generate both images and (noisy) grammars. Let's use that to do some visual scene understanding.
Gudrun
Build an Ackermann robot with RGBD as its primary sense.
Robotics Simulators
Mostly about using Farm Simulator, GTA V, and other games for training both perception and control systems.
Scan for SyncThing Conflicts
Cross-platform Python app to quickly compare conflict files created by SyncThing
Equal Space
Nature Comm. 2022. Use nonlinear manifold learning to discover automatically both the true dimensionality and the underlying spatial coordinates that define a high-dimensional simulation trajectory.
Local Neural Text-to-Speech App
A one-pyfile PyTorch+tk GUI for local neural text-to-speech synthesis.
Faster RNN warmup via manifold learning
Use diffusion maps to skip the warmup phase of RNN inference; demonstrated witha chemical model system.
Learning ODEs from Patchy Observations
Extract Neural ODEs from data whose channels are observed at different times and frequencies.
Certified Invertibility in NNs via MILP
Explore excessive NN invariance in various contexts, with methods for certifying invertibility pointwise across input space.
ANOVA and PCE for Biological Neural Networks
Use ANOVA to perform integrals for polynomial chaos expansions.
Representation Learning
PNAS 2020. Unsupervised learning methods to transform data into a form that's somehow more useful.
Iterative ANNs
Neural networks build on various numerical iterative algorithms.
Learning stochastic DEs from data
Suggest alterative methods for learning stochastic differential equations from data as neural networks.
Hamiltonian Neural Networks
Learn dynamics with constrained quantities.
Gunnar
Build a differential-drive robot with LIDAR as its primary sense.
Meta-learning of ODE integrators
Rather than learning the RHS of an ODE, learn the parameters of the integrator itself.
Learning for Multiphase Flow
After some dimension reduction by PCA and autoencoder, learn an ODE for the slow dynamics of the Navier-Stokes equations in a multiphase flow setup.
QRembed
Embed files into possibly multipart QR codes
Project Opener Menu
A little TK menu for quickly getting to my project directories.
GPT3 for Seminar Announcements
Use OpenAI's API to generate ics files from email text.
Next Task Decider
Process task list and decide what I should do next.
Cat Wrangler
A feline surveillance bot using the guts of an iRobot Braava.
BusinessCardScanner
Split business card images into individual cards and extract info from them.
Budget Monte Carlo
A Monte Carlo simulation Dash app for personal finance, comparing rent vs. buy scenarios and visualizing long-term net worth outcomes under uncertainty.
Boston AV Group Robocar
Teach a one-week workshop to high school students on building and programming a small autonomous car.
Hierarchy Formation
Simulate the formation of dominance hierarchies through social combat.
Circadian Rhythms
Simulate circadian rhythms in the suprachiasmatic nucleus of the hypothalamus.