Anoop Kumar

Indian Institute of Science, Bangalore

GPA: 7.9*

Major Courses: Deep Learning For NLP, Machine Learning, Practical Data Science, Computational Methods of Optimization, Linear Algebra & Probability, Algorithms, Compilers

GCET, Greater Noida

Percentage: 82.26

Major Courses: Operating Systems, Computer Organisation, DBMS, Computer Networks, Data Structures & Algorithms, Engineering Mathematics

Bharatiya Vidya Bhavan's KDK Vidya Mandir, Renukoot

Percentage: 89.0

Data Science Intern

Developed a Machine Learning pipeline to generate Personalized User Recommendations.
- Prepared User session dataset from raw data
- Performed required Exploratory Data Analysis
- Trained embedding models to capture the contextual representation of the products & Users
- Ensemble version of MetaProd2Vec gave a 10 % improvement in mean cosine similarity.

Advance Chatbots

- Working on the System Act Prediction and Natural Language Generation (NLG) modules.
- Developed a GPT based transformer model for Response Generation in Few-shot scenario.
- The model achieved BLEU score of 63% and Slot Error Rate of 3.1%.

Brain Networks

- Modelled human brain as a network using fMRI data, to detect Alzheimer.
- Built a Hierarchical GCN based model, achieved 83% accuracy on ADNI data, beats kernel-based methods.
- Working on scalable multiplex embedding techniques to incorporate multi-task fMRI.

Aspect-Based Sentiment Analysis (ABSA)

- Developed a NLP model to predict aspects and its polarity from restaurant reviews.
- Constructed dummy sentences from the aspect to convert ABSA to a sentence-pair classification task.
- Fine-Tuned the pre-trained model from BERT to achieved 85% F1 score on Test data.

Sentiment Analysis on Code-Mixed Text

- Developed a NLP model for sentiment Analysis on Code-Mixed Hinglish Dataset (Semeval 2020)
- Used an ensemble architecture of convolutional neural net (CNN) and self-attention based LSTM.
- Code-Mixed data is handled using XML-R, achieved 68.5% F1 score on Test data.

Few-Shot Link Prediction in Knowledge Graphs

- Made key modifications to the existing MetaR model for few-shot link prediction
- The new model uses a semantic scoring function and pre-trained model entity embeddings
- The model improves the MRR from 26.1% to 31.5% on NELL-One dataset in 5-shot setting

Prototypical networks for few shot image classification

- Built a deep metric learning based Few-shot classification model
- Primarily a CNN based model implemented using Pytorch. Dataset used: Omniglot.
- The model improves the classification accuracy from 85% to 87.6% in 1-shot setting.

Loop Interchange Tool in MLIR

- Developed a loop interchange tool in LLVM to speed up matrix operations
- The tool optimizes for locality (both spatial & temporal) and parallelism for multi-cores.
- Driven by an analytical cost model, implemented on Affine dialect in MLIR.