Used Multiple Instance Learning to improve disease classification in images spanning millions of pixels.

Implemented an unsupervised learning technique called Contrastive Predictive Coding to create image representations that leverage unlabeled data

Alexa-NLU. Common Sense Reasoning.

Was a member of the Visual Understanding Division where I made the following contributions:

1)Trained deep neural networks for image classification and object detection in Caffe and Tensorflow.

2)Brainstormed and implemented different training techniques for fast and robust on device real time image classification. Developed new methods to reduce training time for a continuously expanding category catalogue.

3) Designed a novel architecture for Human Object Interaction Detection which achieved close to state of the art accuracy on public data sets.

4)Further refined the architecture to reduce model complexity and size. Experiments done in Pytorch and Tensorflow.

Above solutions are commercialized in Bixby Vision employed in S9 Samsung Series, Photo Gallery of S10 and upcoming S11 Samsung Series.

5) Worked with other team members on Salient Moment Detection in Videos, Video Classification, Instance and Semantic Segmentation.

Added street crossing alignment functionality to Autour, an eyes-free mobile system

for the blind using image processing algorithms to identify crossings and orient the

user accordingly. Advised by Dr Jeremy Cooperstock.