Microsoft - Deep Learning

I worked on various projects involving Deep Learning, starting with image acquisition, model training, evaluation, and finally deploying the trained model into the cloud. Based on these projects, I published tutorials for image classification (readme, code), object detection (code+readme), or image similarity (code+readme).

For example, I designed the anomaly detection model used in a live demo during Satya Nadella's Build 2018 keynote.

Another project was with Liebherr, a Swiss equipment manufacturer. The task was to find grocery items in refrigerators. At that time Microsoft's Deep Learning toolkit (CNTK) did not have any support for object recognition. Together with an intern whom I supervised, we added a new DNN layer to CNTK for Fast R-CNN object recognition, with full support for CPU and GPU, forward and backward propagation. More details about this work can be found in this the Azure blog.

Or another project where the goal was to rank clothing items based on the similarity of their texture (dotted, striped, etc), neck (v-neck, round-neck, etc) and sleeve length (no sleeve, long sleeve, etc). For this, every image was represented using the DNN embeddings of a trained model, and by learning a function to compare how similar two images are. Finally, the trained model was published as Rest API with a simple UI for visualizing the ranking results.

During the early days of Deep Learning, I added a DL featurizer module inside the Azure ML Studio product. The module is shown below, and takes images as input and returns the respective ImageNet pre-trained CNN embeddings.