Lyn Walker

Machine Learning Engineer

Matician•  September 2019 - September 2020

- Refined and developed tools and pipelines for data processing, training, validation, and testing of deep learning models. Improved pipeline and tools for unsupervised and semi-supervised training of depth-estimation networks. Developed tools for training, validating, and testing a visual camera pose estimation network with both unsupervised and supervised training modes. - Trained, bench-marked, and analyzed deep learning models to give insights into architecture and pipeline refinement. Improved the accuracy of a deep learning network architecture for estimating changes in camera pose over a stereo video sequence by 10x in both relative and absolute error metrics. Improved the error score of a depth estimation deep learning network from 0.091 to 0.080 on GPU, and from 0.130 to 0.087 on an edge device. - Adapted deep learning models to improve performance on an AI vision processor (edge device) after model quantization and translation, taking into consideration requirements for an autonomous consumer robotics product. Analyzed trade-offs in edge device fps, image size, accuracy, and more to present and compare multiple options for on-device models. - Studied computer vision, deep learning, visual geometry, and machine learning topics in depth to get an advanced understanding of relevant challenges, existing solutions, and directions for research and improvement.