CV - Saurabh Daptardar

Work experience

Xaq Lab: Research Assistant, (2016 - 2018)

Designed experiments to study decision making/control in the animal brain.
Formulated, modeled, and solved the control problem for the designed experiments.
Proposed novel approach to explain and detect change of mind/decision using the control framework.
Developed novel framework for Inverse Reinforcement Learning in POMDPs with continuous states and continuous actions.
Applied the framework to recover the internal latent parameters for the brains’ control model.
Predicted the choices made by monkey’s with 96% accuracy with our model with the recovered parameters.
Predicted change of mind ~100 ms prior to the intended actions using our model.
Culminated the research in my Master’s Thesis “The Science of Mind Reading: New Inverse Optimal Control framework”.

Samsung R&D India: Senior Software Engineer

Samsung Auto Connect: Car and Driver Analytics
- Developed a solution for driver profiling and scoring.
- Trained algorithms for detection and classification of driving maneuvers with 92% accuracy.
- Modeled and developed dynamic context based fuel estimation.
Samsung Gear Smartwatches: Sports Analytics
- Developed application for smart self tutoring for Tennis and Badminton in ‘Gear S2’ device.
- Engineered data collection (from professionals), preprocessing, and storing pipeline.
- Worked on providing smart recommendations, feedback and action matching based on professional players’ strokes.
Samsung Smart Glove (research)
- Designed wearable gloves (with sensors) to control Samsung devices.
- Built gesture recognition and user command interpretation based on sensor signals from the gloves with 82% accuracy.

Projects

Parameter Space Inverse Reinforcement Learning: Master’s Thesis, (2016 - 2018)
Dr. Xaq Pitkow

Developed a novel generalized framework for Inverse Reinforcement Learning to recover true parameters accurately within error bounds.

Control and Reinforcement Learning (RL)

Implemented and trained various control and deep RL algorithms: LQR, Iterative LQG, deep Q Networks (DQN), Double DQN, deep deterministic Policy gradient (DDPG), Advantage Actor Critic (A2C) etc.

Decoding neural activity to estimate control target (Neural Signal Processing)

Decoded reach targets with 84% accuracy using the plan period and movement period signals from the dorsal pre-motor cortex.

Neural decoding: ECOG data to speech (Statistical Learning)

Implemented ensemble methods to classify the neural recordings into dictionary of words with an accuracy of 74%.

Image Captioning with RNNs

Character Sequence RNNs

Implemented and trained a single layer RNN (LSTM) char by char model to generate text.

Single Frame Image Super Resolution: B.Tech Thesis, (2013 - 2014)
Prof. Devendra Jalihal

Implemented the kernel Hebbian algorithm for single frame image super resolution.
Integrated algorithm into a web application for medical/agricultural advisory.

Daptardar, Saurabh, et al. (2015).

CV