Akshita Guptaअक्षिता गुप्ता

I am an ELLIS PhD student at TU Darmstadt, co-supervised by Prof. Marcus Rohrbach and Dr. Federico Tombari at Google Zurich. I completed my MASc at the University of Guelph, where I was advised by Prof. Graham Taylor. During that time, I was also a student researcher at the Vector Institute.

I was fortunate to spent time as an research intern at Apple under Dr. Tatiana Likhomanenko, Microsoft under Gaurav Mittal and Mei Chen, Vector Institute under Dr. David Emerson, and as a scientist in residence at NextAI Prof. Graham Taylor.

Before starting coming to academia, I worked as a Data Scientist at Bayanat, where I focused on projects related to detection and segmentation. Prior to that, I was a Research Engineer at the Inception Institute of Artificial Intelligence (IIAI), working with Dr. Sanath Narayan, Dr. Salman Khan, and Dr. Fahad Shahbaz Khan. At IIAI, my research primarily involved open-world and zero-shot object detection, generative adversarial networks (GANs), and few- and zero-shot learning.

Email  /  Google Scholar  /  Twitter  /  Github  /  Resume/CV

I'm interested in developing models which can learn with limited data and few, zero or one training sample(s).

• Description: Developed multi-scale context aware detection framework with attention-driven psuedo-labelling.
• Outcome: Improved state-of-the-art performances on MS-COCO dataset with absolute gains ranging from 1.8% to 3.3% in terms of unknown recall.

• Description: Developed a attention module which combines both region-level and global-level contextual information.
• Outcome: Improved state-of-the-art performances on NUS-WIDE, OpenImages by 6.9% and 31.9% mAP score.

• Description: Developed a generative model that constructs multi-label features for (generalized) zero-shot learning.
• Outcome: Improved state-of-the-art performances on NUS-WIDE, OpenImages and MS-COCO by 3.3%, 4.3% and 15.7% mAP score.

• Description: Developed a generative feature synthesizing framework for zero-shot learning.
• Outcome: Improved state-of-the-art performances on CUB, FLO, SUN, and AWA by 4.6%, 7.1%, 1.7%, and 3.1% harmonic mean by enforcing semantic consistency at all stages of zero-shot learning.

• Description: Improved state of the art object detector (Mask-RCNN and PANet) for aerial imagery.
• Outcome: Proposed a large scale instance segmentation and object detection dataset (iSAID) with benchmarking on mask-RCNN and PANet.

• Description: Developed an attention based framework for acoustic scene recognition and audio tagging.
• Outcome: Improved the equal error rate by atleast 3% over the Dcase challenge results.

• Developing deep learning algorithms for low- (Few- and zero-) shot detection and classification, generative adversarial models and open-world object detection problems.
• Developed rock & seismic layer classification system.
• Worked on satellite-imagery object detection and object counting system.

Developed an open source analytics dashboard prototype with the metrics to evaluate diversity and inclusion across different communities.

Worked on acoustic scene recognition and audio tagging using attention networks. Paper accepted in Interspeech-CHIME 2018.

I borrowed this website layout from here!