Balakrishnan Varadarajan

- Develop efficient image classifiers for blocking bad content.

- Define metrics, data pipeline for human annotations.

- Refine generative AI outputs based on human feedback to make them safe and reliable for public use.

- Develop efficient vision models for ondevice image understanding

• Developed novel approaches for 3D localization of objects via spatio temporal fusion.
• Bring state of the art multimodal AI models to AR/VR devices under constrained enviroments.
• Worked on prototyping multimodal AI technologies for future AR/VR devices.

Built state of the art behavior prediction systems for predicting the future trajectories of neighboring agents.

• Deep learning for behavior prediction: I spearheaded several key architecture designs that would enable deploying larger deeper models on the car. This shift to deep learning for behavior prediction enabled us to rapidly iterate on models and architecture and this led to the behavior being a order of magnitude safer around areas with high pedestrian density.
• Academic contributions: I have collaborated on three publications with research and perception teams. Behavior prediction is a rapidly evolving field and the contributions are valuable to the community. Our latest publication (multipath++) is published in arxiv as well as accepted in a top conference (ICRA). Our approach gets state of the art prediction results on atleast two external behavior prediction benchmark datasets.

Applied research: I worked on deep learning for video understanding with focus on products (products include YouTube, Google photos and Cloud). Our primary focus is achieving human level performance wrt understanding of video content (pixel, audio and metadata).
We also build custom multimodal machine learning models for various YouTube products (ads, brand safety, recommendation) going from the raw video content.

Over the past few years, the information achievable from the raw video content has surpassed the one that can be obtained from metadata. Me and my team are the primary drivers of this technology.

We also developed novel techniques to scale to large amounts of data (videos in particular). Our work was published in top conferences like CVPR.