I currently work at Ederlabs on privacy preserving Machine learning (PPML).
I am also a member of [Openmined](https://www.openmined.org/) SyMPC team.
My interests are in Machine Learning, Privacy Preserving Systems and Software Engineering.
In PPML , I am specifically interested in Differential Privacy, Federated Learning and Trusted Execution Environments (TEE's).
Available on Twitter as kamathhrishi and on LinkedIn.
I would like to write Blogposts
Experience
-
Privacy AI Engineer , Ederlabs (October 2020-Present)
Working on privacy preserving machine learning using Trusted Execution Environment(TEE) Intel SGX.
-
Intern , Ederlabs (January 2020-September 2020)
Worked on Differential Privacy and Intel SGX
-
Freelance Web Developer (July-October 2018)
Worked on two frontend projects for two clients using AngularJS and Bootstrap4 .
Open Source Contributions
PySyft
“A library for answering questions using data you cannot see”
A framework for privacy-preserving deep learning using Pytorch and Tensorflow.
I worked on Pytorch development.
Part of the Core team
Contributed to Core PySyft Team
The experience helped me learn a lot of software engineering practices such as unit tests, code reviews, Git, writing clean and documented code.
- Wrote use cases of Federated Learning for developing Word Embeddings from Private Data and demonstration on CIFAR10
- Wrote an implementation of Differential Privacy method Private Aggregation of Teacher Ensembles (PATE)
- Tried working on Polynomial Tensor for non-linear computation in multiparty computation (MPC) setting. The method approximated non-linear computation using interpolation/Taylor series methods. Could not integrate it as part of Syft chain of tensors.
- Refactored data loaders and federated dataset and wrote a tutorial on developing custom federated datasets
- Type annotated the codebase and wrote documentation for parts of it
Deepgaze
Computer Vision library for human-computer interaction in Python
I have refactored certain parts of the library for improved usability and wrote test cases to ensure accurate functionality. Ported library to Python 3.0.
Some Projects I am proud of
For complete list of projects feel free to check my Github Profile.
GreyNSights
Privacy Preserving Pandas
Indoor Scene Recognition with Visual Attributes
- Trained a Residual Network to classify 67 different Indoor Scenes on MITIndoor 67, obtained 60% accuracy on the test dataset.
- MITIndoor67 is a small dataset consisting of 5400 images, training a network directly on the dataset led to substantial overfitting.
- In order to learn finer features, a separate attribute prediction network was trained on SUN attribute database.
- The predicted attributes are used to augment image features in the linear layers of the network scene recognition network.
Weakly Supervised Street View Text Detection
- Used Pytorch, OpenCV and Pillow in Python
- Trained a character agnostic text detector on Chars74K dataset along with images consisting of indoor/outdoor scenes without text. The character agnostic model is a alexnet network pretrained on imagenet.
- Used the classifier and sliding windows to annotate images in UCSD SVT and NEOCR dataset to derive bounding boxes
- Trained a street text localisation and detection Fully Convolutional Network(FCN) on the weakly supervised labelled dataset by using text detector’s starting CNN layers
- Currently working on character level segmentation and recognition
Animals with Attributes
Supervised Learning
- Trained a 30 layer Residual CNN from scratch which could predict 50 different species of animals with 72% accuracy. Trained it with probabilistic augmentations using Alubmentations library.
- Finetuned an ImageNet pre-trained 18 layer network to attain 89% accuracy
- Spent time understanding and analyzing top Imagenet CNN architectures
Zero Shot Learning
- Currently exploring Zero shot learning on the dataset. A naive implementation with Direct Attribute prediction attains 35% accuracy prediction unseen classes.
Contact me on kamathhrishi [at] gmail