Brian McCrindle
About Me
Hey, I’m Brian! I specialize in computer vision, physics-informed machine learning, and optical systems. I come from a different background than most when it comes to computer science, but I think it adds a nice “spice” to a traditional ML team. :)
Before getting into machine learning, I studied Engineering Physics at McMaster University where I focused on lasers, electro-optics, and “old school” computer vision.
My Master of Applied Science was in the department of Electrical and Computer Engineering focused on improving uncertainty quantification for medical computer vision systems under Dr. Michael Noseworthy. I was also a MITACS scholar during this time, working on physics-informed machine learning with Solid State AI to improve the growth of semiconductor-based lasers. After graduation, I continued to work with Solid State AI to work on building Solid State’s deep learning toolbox, and create their anomaly detection product for manufacturing applications.
Currently, I am a machine learning research engineer at Macrocosmos AI, an group working on decentralized AI on the Bittensor protocol built by the Opentensor Foundation.
Previously, I worked as a machine learning engineer at Parallel Domain. PD creates image-based synthetic data to improve model performance for perception tasks, primarily for automotive and drone applications. Here, I conducted research on how to effectively use synthetic data for various tasks (2D/3D detection, 2D/3D semantic segmentation, optical flow, depth estimation, etc..), generative modelling to increase asset diversity for better model generalization in the long-tail, engage and guide with our customers with these insights, and help develop our ML research tooling.
You can refer to my linkedin for additional information.
Highlights
CVPR 2023: On the Importance of Label Domain Gaps - Phillip Thomas (joint work at Parallel Domain)
The real-to-synthetic domain gap is a massive problem when wanting to leverage synthetic data during training. If this gap isn’t closed, it is highly likely that feature representations learned during synthetic pre-training will be forgotten (also known as catestrophic forgetting). In this talk, Phillip goes through some preliminary research we did at Parallel Domain to uncover the importance of matching label specifications between the source and target distributions.
McMaster Engineering Capstone Highlight 2019: NanoRIMS:
NanoRIMS is a benchtop lab system to automate the synthesis of gold nanoparticles for biomedical sensor research. As a team of 4, my sub-tasks were focused on creating a purpose-build spectrophotometer to measure particle size while still in solution, and estimating AuCl3 droplet mass from the nozzel output (on the order of micro-grams). This work was done in collaboration with Dr Leyla Soleymani to reduce her lab’s downtime.