Greater Toronto Area, Canada
- Investigated the utility of state-of-the-art (SOTA) Siamese self-supervised learning (SSL) methods (SimCLR, Barlow Twins, and VICReg) for the unsupervised pre-training of M-mode lung ultrasound images - Created specializations of SOTA SSL methods to leverage the unique qualities of the M-mode ultrasound data domain
- Designed, implemented, and evaluated neural networks for the classification of the lung sliding artefact in lung ultrasound videos - Implemented video classification architectures, including temporally distributed CNNs and inflated 3D CNNs, and image classification architectures, such as EfficientNets and MobileNets, using Tensorflow - Leveraged Tensorboard for training and evaluation metric visualization - Implemented Grad-CAM to generate feature importance heatmaps - Implemented a configurable input pipeline for lung ultrasound videos and transformed images using Tensorflow Datasets
- Applied isolation forests to tabular call data to detect anomalous customer calls, helping to identify impactful improvements in call handling procedures - Experimented with feature engineering and hyperparameter searches using sklearn
- Prototyped various RetinaNet-based models in PyTorch, using attention condenser and standard convolutional backbones, for PCB component detection in the FICS PCB dataset - Implemented test-time augmentation, for classification and object detection tasks - Built data pipelines for object detection tasks for data stored in COCO and Pascal VOC formats, including refactoring of existing utilities and old pipelines to decrease redundancy