Resume¶
Dan G. Jacobellis | danjacobellis@utexas.edu
Education¶
Pursuing PhD in Electrical Engineering • University of Texas at Austin • 2021-Present • GPA: 3.65
Notable Coursework:
Information Theory
Statistical Machine Learning
Physical Acoustics
Spoken Language Technologies
Digital Video Processing
Parallel Algorithms
Computational Methods for Inverse Problems
Advanced Computer Vision
M.S. Electrical Engineering • University of Texas at Austin • 2018-2021 • GPA: 3.67
B.S. Electrical Engineering • University of Texas at Austin • 2013-2018 • GPA: 3.71
Awards¶
2022 and 2023 Top Student Teaching awards, ECE Department¶
Awarded in recognition of excellence in teaching as a teaching assistant for ECE 445S Real-time Digital Signal Processing Lab.
2019 Research Excellence Award, Applied Research Laboratories¶
Annual award in recognition of outstanding research efforts based on nomination from peers and colleagues.
Experience¶
ECE Department, University of Texas at Austin¶
Graduate Research Assistant • 2021 – Present
Graduate Teaching Assistant • 2019 – Present
I am currently a graduate research assistant in the UT SysML lab, where I research and develop machine-oriented compression systems.
From Fall 2019 to Fall 2023, I was a teaching assistant for ECE 445S Digital Signal Processing Lab. I led the laboratory section of the class where components of a software defined radio system are implemented.
In addition to my TA position, I was appointed as a GRA to develop new course materials. When courses were moved online due to COVID-19, I assisted in restructuring the lab work so that it could be completed online using the audio hardware built into students’ personal computers. With the return to in-person classes is 2021, I developed a new lab manual for the course to help transition to new hardware. My contributions include
Development of starter code and tutorials for programming the Cortex-M7 based STM32 development board
New, concise explanations of DSP topics including
Dozens of new figures, illustrations, and code examples
Development of new lab exercises, including an acoustic modem and vocoder
Tutorials for using the ARM CMSIS DSP libraries
Other TA positions:
ECE 382V Systems and Machine Learning (Spring 2024)
ECE 351M Digital Signal Processing (Spring 2023)
ECE 351K Probability and Statistics (Summer 2022 and Fall 2022)
ECE 313 Linear Systems and Systems (Summer 2019)
Modern Intelligence, Austin, TX¶
AI Research Scientist Intern • May 2023 - Nov 2023
Lead researcher on several projects:
Representation learning for multichannel acoustic, radio, and hyperspectral signals.
Split computing for low power and low bandwidth remote sensing.
Generative signal enhancement with unknown corruption operators
Applied Research Laboratories, University of Texas at Austin¶
Graduate Research Assistant • 2018 – 2022
Research Engineering Scientist Associate • 2017 – 2018
Student Technician • 2015 - 2017
Developed software and performed underwater acoustics research for the Environmental Science Laboratory. Work included:
Development of MATLAB software tools to perform geoacoustic inversion
Analysis of newly collected data from passive acoustic sensor arrays
Software performance optimization for large computing clusters
Researching new methods to quantify uncertainty of ocean acoustic propagation models
Projects¶
Learned Compression for Compressed Learning
Developed WaLLoC (Wavelet Learned Lossy Compression), a framework for efficient machine-oriented compression that is compatible with a wide range of modalities (audio, images, etc). WaLLoC provides comparable quality to generative autoencoders like the VAE in Stable Diffusion, but provides (1) efficient encoding (>1 megapixels per second on raspberry pi) (2) greater dimensionality reduction (16x vs 12x in SD3) and (3) high compression ratios (35:1 vs 6:1 in SD3).
-
Split computing is often used for wearable sensing (e.g. smart glasses) to offload expensive ML models to the cloud. By adapting a neural codec to the specific sensor, extreme compression ($>$500:1) can be achieved while retaining very low complexity (~500 MACs/pixel).
-
Evaluated several popular models—including image classification, image segmentation, speech recognition, and music source separation—under severe lossy compression. Results establish several key findings for evaluating lossy compression in the context of machine perception pipelines.
-
Current approaches that adapt diffusion models for audio discard phase and require a vocoder to compensate. The MDCT (the same time-frequency transform used in MP3) can be used as a replacement in generative diffusion models.
-
Developed a new class of perfect-reconstruction filter banks for audio analysis. It uses a novel type of time-frequency tiling to retain the benefits of constant-Q transforms while retaining an efficient uniform-grid spacing for each sub-band. Implemented on GPU via CUDA.
GPU accelerated non-negative matrix factorization for audio
Implemented multiplicative update matrix factorization algorithm in CUDA for applications of audio source separtion. Performance on consumer GPU outperforms single-core Scikit-learn/OpenBLAS by a factor of more than 100.
Skills and expertise¶
Programming
Pytorch
Julia
C
MATLAB
Signal Processing
Time-frequency analysis
Wavelets, filter banks
Denoising
Compression
Machine Learning
Representation learning
Neural compression
Self-supervised learning
Generative models
Acoustics
Phased array processing
Inverse problems
Propagation models
Miscellaneous
Cooking
Cello
Sewing