My upstairs is a chaotic garden of ideas, some meh, some decent. I lucked out being born and raised in India in the late 80s and 90s, where there was no such thing as psycho-analysis of children. I am pretty darn sure that I would have been boxed in some corner of the spectrum given my propensity to run amok and frolic between many ideas. I find myself working on half a dozen projects at any given time and here's a list below. I am currently working on a few of these under the aegis of the UnifyID AI fellowship where I mentor a cohort of 9-12 fellows every semester. HMU if any of this intrigues you or better yet, apply to the fellowship if you live in the bay area.

  1. Is augmentation all you need?
  2. Art attack: Using artistic style transfer to generate adversarial perturbations
  3. Gibberish-in-Gibberish out? How do SoTA Language-ID classifiers handle human generated gibberish?
  4. Emotive speak: Building an agent that does emotion detection and converses using on device ML with the Google AIY kit
  5. SoTA assessment of real world susceptibility of adversarial attacks n the Google AIY kit and Amazon DeepLens
  6. Effect of mixed precision inference on adversarial attack susceptibility
  7. Bandpass filters in momentum specification in Gradient descent ideas
  8. Lautum information bottleneck
  9. Real world 'explainable' one-shot learning on three 'hard' novel datasets
  10. Homophonic puns and NLP: Wavenet and beyond
  11. On celeb identification and doppelgangers
  12. Explainable AI: GRF and features from pre-softmax layers
  13. RogueNets - resnets with long range connections
  14. Multi-step training for training Gait classification models
  15. Usage of style transfer for data augmentation and soft labels
  16. Between-class Learning for Gait classification
  17. Are shallow trainable examples of a training set also one-epoch trainable?
  18. Change point detection for human motion data
  19. On the occurence of 'Somewhat surprisingly' in ml papers
  20. Training StackGaNs with hashtags (Hashtag2Image)
  21. Adversarial perturbations and augmentation
  22. On optical illusions and Mask-R-CNN
  23. Generative models for human-motion time-series data - GANs
  24. Human-motion time-series data classification
  25. Flow-based generative models for human-motion time-series data
  26. Explainable AI in Time-series data: Visualizing what the CNN classifier 'saw' in time-series classification?
  27. On Tsallis and Renyi metrics: Non-shannonian loss functions and CNNs
  28. Loss landscapes, adversarial augmentation and training on randomized labels: An exploration
  29. Loss landscapes and the Information plane visualization
  30. Hashtags and word embeddings
  31. Open source EE predictor using commercial smartphones
  32. Quaternion-neural-networks for gait classification
  33. Deep Universal Background Models
  34. Heels and Gait
  35. Sitting posture measurement and analysis using IMU sensors
  36. Squatting and IMU : Proper form and squatting deviation classification
  37. Softmax bottlenecks and outrageously large class classification
  38. Exoplanet Hunt with the Kepler data