
I am a machine learning researcher interested in deep generative models, neural representation learning, reinforcement learning, and AI for climate change. I am a postdoc at the National Renewable Energy Lab (NREL) where I develop machine learning algorithms for problems such as building energy management, weather and power forecasting, and discovering novel enzymes. NREL is a DOE national lab that uses AI and supercomputing to help tackle the climate crises, which I believe to be one of the most pressing challenges of our time. I recently graduated with my Ph.D. in Computer Science from the University of Florida in December 2021. My Ph.D. research was on neural scene understanding and spanned topics including: object-centric deep generative modeling, dynamic point cloud modeling, and low-resource multi-object tracking for traffic signal control.
[CV][Google scholar] [Github]
Contact: pemami[at]nrel[dot][gov]
News
- December 2022 I will be in New Orleans for NeurIPS’22 (Nov 30th-Dec 3rd). Come say hi!
- October 2022 One paper accepted at the Machine Learning for Structural Biology workshop at NeurIPS’22!
- April 2022 Recognized as a Highlighted Reviewer at ICLR’22!
- January 2022 I am now a postdoc at NREL!
- December 2021 I defended my thesis “Neural Algorithms for Object-centric Scene Understanding” and graduated with my Ph.D.!
- December 2021 One paper accepted at AAAI’22 (15% acceptance rate)!
Recent Papers
Plug & Play Directed Evolution of Proteins with Gradient-based Discrete MCMC
A fast MCMC sampler for discovering variants by mixing and matching unsupervised evolutionary sequence models with supervised models that map sequence to protein function.
Learning Scene Dynamics from Point Cloud Sequences
We introduce the sequential scene flow estimation problem and provide a novel baseline architecture and benchmark for evaluation.
Long-range Multi-Object Tracking at Traffic Intersections on Low-Power Devices
A CNN architecture for joint object detection and Re-ID on low-power edge devices and a multi-sensor tracking algorithm for radar.
Efficient Iterative Amortized Inference for Learning Symmetric and Disentangled Multi-Object Representations
An efficient slot-based hierachical VAE for learning unordered, symmetric, and disentangled scene representations.
A Symmetric and Object-Centric World Model for Stochastic Environments
We propose a state space model and training objective that achieves better stochastic future prediction for object-centric world models.
See my CV for a complete list including older publications.

Leeuwenberg, E. L. J. "A Perceptual Coding Language for Visual and Auditory Patterns." The American Journal of Psychology, vol. 84, no. 3, 1971, pg. 338.