Hi — I’m Sutirtha, a Machine Learning Researcher. My interests broadly lie in Interpretability in Deep Learning Architectures. I am also currently exploring the domain of Continual Learning to adress the Catasthropic Forgetting Problem.
I have worked as a Machine Learning Engineer at a healthcare startup, where I have dealt with several Computer Vision related problems, such as Multi-task Learning, Class-Imbalance Problem and Image Segmentation Problem.
I have been an Undergraduate Researcher at Jadavpur University, where I have worked on domains of Machine Learning in Medicine and Agriculture. I have also worked on modelling Long-Time Series Forecasting sequential data.
- Interpretable and Prunable Neural Networks
- Continual Learning
- Self-Supervised Learning and World Models
- Multimodal Architectures
- Long Time Series Forecasting
The following organizations represent the anthologies of my research experiments:
-
Interpretable Nerual Networks — This organization contains repositories, that analyses the low-level information and gradient flow in the neural networks. Especially, low-level gradients and neural bottlenecks have been observed.
-
Continual Learning — This organization contains repositories, trying to observe the Catastrophic forgetting problem and the ability of Continual and Incremental Learning to adress this problem.
- This repository represents my first venture to research. This work has been accepted by Willey (Currently under Final Publication Process). [Github Repo: (https://github.com/SutirthaMukherjee97/PlantLeafDetection)]
- The general requirement of this work is to develop a classification model, that can detect the particular plant species, based upon the given input image of its leaf. The specific object of this work is develop a Model with low training data-samples, that is light-weight, well generalized and is highly accurate over unseen test samples.
- The methodology involves training a DL Model (InceptionResNetV2) and leveraging the HSV and YUV color spaces, apart from the RGB, which generates three sets of confidence scores for that image. The ensemble of the of the three outputs have been done using Sugeno Fuzzy Ensemble scheme.
- This work has been accepted at the CPAMCS-2023 conference, hosted by Springer, 2023. [DOI: https://doi.org/10.1007/978-3-031-64010-0_37]. [Github Repo: (https://github.com/SutirthaMukherjee97/PneumoniaDetection_CXR_Images)]
- The high-level objective of this work is to develop a DL Model system, that can reliably classify Pneumonia, COVID-19 and Normal Images. The specific objective is to develop a learnable feature selection algorithm over a generalized model system, to produce highly accurate prediction scores.
- The architecture involves training separately a ResNet50 and a ViT architecture. It is followed with feature reduction using PCA. Finally, these features are meta-selected using XGB model, to produce the final output.
- TBA
- Email: sutirthamukherjee1@gmail.com
- Google Scholar: https://scholar.google.com/citations?user=95-I7HoAAAAJ&hl=en
- LinkedIn: https://www.linkedin.com/in/sutirtha-mukherjee/
- Twitter: https://x.com/StoryTeller_97