Machine Learning

Our industrial experience is primarily in the field of computer vision, specifically for solving problems related to image segmentation, classification, and object tracking. 

• We've worked on several projects involving binary and multiclass classification using standard architectures such as ResNet, EfficientNet, and others. One of the more interesting projects involved the classification of a large number of classes (>10,000) with the ability to classify instances not seen during training but only in the production environment. This problem was addressed by using metric learning, mapping images to embedding space, and applying nearest neighbor based classification of images in the production environment. 

• For object tracking, we used a tracker based on the deep sort paradigm, which relies on detections obtained from standard object detectors such as RCNN and YOLO, motion estimates of these detections obtained using the Kalman filter, and visual similarity estimates of detections based on their embeddings.

• In one instance segmentation project, we performed the segmentation of cell nuclei and cytoplasm. Since the dataset was very limited, we fine-tuned open-source models such as StarDist2D, CellPose, and HoVer-Net. 

• The problem of autofocus in electronic microscopes is one of the critical issues in bioimaging that researchers continue to explore actively. We automated this problem by using a neural network to predict the distance from the focus based on the various focal measures. Such estimates were used in a custom made optimization algorithm for considerably faster autofocusing than traditional microscope focusing methods. 

 

In addition to the mentioned experience, we have industrial experience in data handling (images), data annotations (planning, management, and validation), MLOps, and optimization and deployment of models. 

 

Apart from the mentioned industrial experience, we also have academic research experience, primarily in graph-based learning, few-shot learning, fairness in machine learning, and modeling of spatial data. We have published papers in top journals and conferences, such as NeurIPS and ICML.