About me

  • I graduated at the top of my class with B.Tech in Electrical Engineering from Indian Institute of Technology Madras and and currently pursuing my Masters degree from Stanford University.
  • My research and industrial interest lay in computer vision, reinforcement learning, and using AI for health care or other social goods. I am also interested in understanding and mitigating failure modes that limit the robustness and reliability of Neural networks.
  • I have had the privilege to intern with Walmart Labs, Google Research, Microsoft and with BridgeI2I Analaytics Solutions during my undergraduate days.
  • My works have also been published in top conferences and journals like COLING, ACL, and Elsevier's Building and Energy Journal.
Experience

Work Experience

Walmart Labs, India - Data Scientist Intern Feb - Jul '22

    Project Title : Medication Scanning – Automatic Retrieval of information from medicine bottles
  • From a given set of bottle images, we automatically retrieve required information about patients to remind them about refills and discount/purchase options.
  • Built a pipeline comprising of blur detection, saliency detection, image segmentation (Mask-RCNN based label localization), and image stitching modules. This stitched image is then passed through OCR and Entity recognition modules to retrieve the required information.

Google Research India - Research Associate Intern Jul - Dec '21

    Project Title : Evading Simplicity Bias in Neural Networks
  • Trained neural network models on a Novel redundancy loss (inspired from self-supervised learning concepts) to make them focus on learning complex features (to overcome Simplcity Bias) and not just on easily learnable parameters and spurious correlations present in the data.
  • The nueral network trained on redundancy loss was found to have learnt better features and also improved the Out of distribution (OOD) accuracy by 5% and Few-shot accuracy by 12% on average from the baseline for a RESNET-18 model on Patch-CIFAR dataset.

Microsoft India - Software Engineer InternMay - Jul '21

    Project Title : Shared Disk Data Tracking for a failover cluster
  • Worked on both Control and data plane to report the shared disk and all IOs in context of owner node.
  • Wrote a script to automatically detach the shared disk from the owner node and reattach the disk to all the cluster nodes before failovering to the target side.

BRIDGEI2I Analytics Solutions - Data Analyst InternMay - Jul '20

    Project Title: Neural Embedding and Bi-Partite Graph based Recommender system
  • Built a pipeline of recommender systems comprising of Popularity, clustering, Item-Item association, Bi-Partite graph and Neural Embedding based recommenders.
  • The Neural Embedding based recommender was built specially for handling sparse input data and the graph based association recommender for considering co-occurrences and higher order proximities among items.
  • This new pipeline increased the hits on the recommended products by 8%. You can visit the Project website here.
Projects

Research Projects

Optimal Compression of Convolutional Neural Network for Edge DevicesJan - Jun '22

  • We perform optimal compression of a CNN network used for image classification by identifying the significant and insignificant layers of the network and compressing those using different techniques.
  • We make use of Two Nearest Neighbors (TwoNN) method to identify the significant and insignificant layers from the image classification model.
  • We also propose another model for memory constrained settings (very low memory capabilities) making use of Mixed Precision network topology.

Input Specific Attention Subnetworks for Adversarial Detection Jan - Jul '21

  • This work demonstrates an altogether different utility of attention heads. We build a novel adversarial detection model based on Self-attention heads of the Transformer model.
  • The Input specific attention subnetworks were used for extracting the features used to discriminate between authentic and adversarial inputs.
  • This work was done under the guidance of Prof.Pratyush Kumar Panda and Prof.Mitesh Khapra. This work has been published in Findings of ACL 2022 conference.

Transductive Transfer learning based LSTM-CNN model for Thermal comfort predictionAug - Dec '20

  • A transfer learning based model is presented for accurate thermal comfort prediction in buildings with limited modeling data across different climate zones.
  • Experiments were performed analyzing the impact of individual layers in the architecture, location for optimal parameter transfer and size of the target dataset on model performance to demonstrate the effectiveness and applicability of the proposed transfer learning model.
  • This work has been published in Elsevier – Building and Environment Journal. You can find the Project's website here. This work was done at SEIL Labratory IIT Bombay, under the guidance of Prof. Anupama Kowli, Dr.Nivethitha Somu and Prof.Krithi Ramamritham.

Gesture recognition from swipe keyboard for Indic languagesJan - Jun '20

  • This project aimed at developing a keyboard that supports gesture typing in Indic languages on mobile devices.
  • A Gesture Decoding model that uses a multi-headed Transformer along with LSTM layers for coordinate sequence encoding and a character-level LSTM model for character sequence decoding.
  • A Contrastive Transliteration correction model that uses position-aware character embeddings to measure word proximities and correct spellings of transliterated words.
  • This work has been published in COLING Conference. You can find the Project's website here. This work was done under the guidance of Prof. Pratyush Kumar Panda and Prof. Mitesh Khapra

Academic Projects

Lightweight CNN model for Music Instrument Classification

  • We built a Lightweight-CNN model to classify musical instruments. We compute the Mel-spectrogram features from input audio data. These features are used as inputs by the CNN model.
  • We use a novel data augmentation technique based on the Cut-Mix Algorithm to add robustness to the model.
  • We also analyze the inputs by generating the gradient based Class activation maps (Grad-CAM) to identify class-wise important DCT coefficients from the input audio.
  • This work was part of a Course project from EE5180 (Intro to ML) course under the guidance of Prof.Avishek Chatterjee. You can find the Project's website here.

Green path prediction based on Air quality data

  • Used Time series analysis to forecast the air quality data of a particular region (surroundings of IIT Madras campus) for a period user's choice.
  • Analyzed the variation in concentration of various pollutants during the day and modeled an algorithm to predict the safest travel path (Green path) between the start and destination in terms of best Air Quality (minimal exposure to pollutants).
  • This project was done as a part the Electronics Club at the Center for Innovation(CFI) of IIT Madras. You can find the Project's website here.

Hackathons

Scalathon: Build an Automatic Headline and Sentiment Generator

  • Objective : Process digital content like emails, articles, reports, videos, tweets etc. The task is further broken down into:
    - Theme Identification of tweets and articles.
    - Headline generation for articles which follow the mobile technology theme.
    - If the identified theme is mobile tech, assign a sentiment against the brand described.
  • Even though Transformer models are the current SOTA in language understanding we propose to use a variant of Kim-CNN to take a phoneme input and predict the theme. Further, this method is computationally efficient and showcases faster inference speeds.
  • mBART, a transformer based model from facebook is used for headline generation. It’s an encoder-decoder model which was pretrained with the objective of denoising multiple languages simultaneously.
  • The Non-english tweets are translated to English using the mBart model. We identify mobile brands entities using NER (Named Entity Recognition). We collected an exhaustive list of mobile brands and Aspect based Sentiment Analysis (ABSA) is used to identify a sentiment towards each aspect in the text.
  • This solution secured us the Gold medal at the 9th Inter IIT Tech Meet, organized by IIT Guwahati.
Personal

personal

Article about me

Here is the link to an article written about my work in my College student magazine.

Cricket

I'm a huge cricket fan, and play as an all-rounder for my university cricket team. I represented my state (Tamil Nadu) cricket team at the U-14 and U-16 levels. I have also had the opportunity to captain my college cricket team .

Public Speaking

I was the head of the Oratory club and Oratory contingent of IIT Madras for the year 2020-21. I am an active part of the Oratory contingent since my freshman year and I love speaking in front of a crowd.

Quizzing

I was an integral part of my school Quizzing team. I was part of finals of three national level quizzes.I was a Finalist in TIMES NIE Quiz, TIMES SCIENCE Quiz, Bournvita Quiz Contest to name a few.

Get in Touch

Contact

sanirudh [at] stanford [dot] edu (Stanford)

ee18b073 [at] smail [dot] iitm [dot] ac [dot] in (IIT Madras)

anirudhsriram30799 [at] gmail [dot] com (Personal)

+91 9789898810

Department of Electrical Engineering,
Indian Institute of Technology Madras,
Chennai,Tamil Nadu