Sai Mitheran Jagadesh Kumar

Sai Mitheran Jagadesh Kumar

Princeton, New Jersey, United States
4K followers 500+ connections

Join to view profile

About

I am currently a part of the Machine Learning team at Latent AI, Inc. after graduating…

Activity

Join now to see all activity

Experience

  • Latent AI, Inc. Graphic

    Latent AI, Inc.

    New Jersey, United States

  • -

    Pittsburgh, Pennsylvania, United States

  • -

    Pittsburgh, Pennsylvania, United States

  • -

    Princeton, New Jersey, United States

  • -

    Pittsburgh, Pennsylvania, United States

  • -

    Pittsburgh, Pennsylvania, United States

  • -

  • -

  • -

    Singapore

  • -

    Saarbrücken, Saarland, Germany

  • -

  • -

  • -

  • -

  • -

    Tiruchirappalli, Tamil Nadu, India

  • -

  • -

  • -

    Tiruchirappalli, Tamil Nadu, India

  • -

  • -

  • -

    Coimbatore, Tamil Nadu, India

  • -

Education

  • Carnegie Mellon University Graphic

    Carnegie Mellon University

    4.0/4.0

    -

    Activities and Societies: Graduate Teaching Assistant (18-794, Pattern Recognition Theory; 18-290, Signals and Systems), Graduate Research Assistant (CyLab, AirLab)

    Coursework:

    Fall 2022:
    - Applied Stochastic Processes for AI and ML
    - Machine Learning for Signal Processing
    - Building Reliable Distributed Systems
    - Introduction to Graduate Studies

    Spring 2023:
    - Advanced Deep Learning
    - Visual Learning and Recognition
    - Optimization
    - Career and Professional Development for Engineering Masters Students

    Fall 2023:
    - Embedded Deep Learning
    - Speech Recognition and Understanding
    - Data, Inference and…

    Coursework:

    Fall 2022:
    - Applied Stochastic Processes for AI and ML
    - Machine Learning for Signal Processing
    - Building Reliable Distributed Systems
    - Introduction to Graduate Studies

    Spring 2023:
    - Advanced Deep Learning
    - Visual Learning and Recognition
    - Optimization
    - Career and Professional Development for Engineering Masters Students

    Fall 2023:
    - Embedded Deep Learning
    - Speech Recognition and Understanding
    - Data, Inference and Applied Machine Learning
    - Research (MS Graduate Project)

  • -

    -

  • -

    -

Licenses & Certifications

Volunteer Experience

  • IEEE Graphic

    Reviewer

    IEEE

    - Present 2 years 8 months

    Education

    Reviewer for IEEE Transactions on Neural Networks and Learning Systems (Impact Factor: 14.255)

  • Community Volunteer

    Hope, NIT Trichy

    - 2 years 7 months

    Health

  • Community Volunteer

    Hope Organisation

    - 1 year

    Civil Rights and Social Action

  • U&I Trust Graphic

    Teaching Volunteer

    U&I Trust

    - 7 months

    Education

  • ICLR Graphic

    Student Volunteer

    ICLR

    - 3 months

    Science and Technology

    Selected as a Student Volunteer for the 9th International Conference on Learning Representations, that was conducted virtually for 2021. Contributed to the tech team to resolve issues and stress-test the platform.

    Certificate - https://drive.google.com/file/d/1MBHV4x9ERS6WLKsL1Sbh7jCbQPMw9krJ/view?usp=drivesdk

  • Co-Founder

    GAIA Organization

    - 2 years 3 months

    Environment

    • Planted a new forest called the 'Miyawaki Forest' in the unused barren land of a school campus and maintained it regularly as the saplings came to life, to promote greenery and reduce pollution
    • Visited 'Snehalaya for Animals' and donated the funds raised to help provide shelter to homeless stray animals, associating with 'Snehalaya Dog Shelter' to carry out the needful.
    • Organised and carried out waste collection campaigns in several parts of the city, promoting Swachh Bharat and…

    • Planted a new forest called the 'Miyawaki Forest' in the unused barren land of a school campus and maintained it regularly as the saplings came to life, to promote greenery and reduce pollution
    • Visited 'Snehalaya for Animals' and donated the funds raised to help provide shelter to homeless stray animals, associating with 'Snehalaya Dog Shelter' to carry out the needful.
    • Organised and carried out waste collection campaigns in several parts of the city, promoting Swachh Bharat and raised funds to donate to rural schools and homes for education.

  • Student Volunteer

    ICML

    - 2 months

    Education

    Selected as a Student Volunteer for the Thirty-eighth International Conference on Machine Learning, that was conducted virtually for 2021.

  • Tennis Lead and Mentor

    National Sports Organisation

    - 3 years 9 months

    Health

Publications

  • Rethinking Feature Extraction: Gradient-Based Localized Feature Extraction for End-To-End Surgical Downstream Tasks

    IEEE International Conference on Robotics and Automation (2023)

    Previously accepted at IEEE RA-L, also accepted for publication and presentation at IEEE ICRA 2023.

    Several approaches have been introduced to understand surgical scenes through downstream tasks like captioning and surgical scene graph generation. However, most of them heavily rely on an independent object detector and region-based feature extractor. Encompassing computationally expensive detection and feature extraction models, these multi-stage methods suffer from slow inference…

    Previously accepted at IEEE RA-L, also accepted for publication and presentation at IEEE ICRA 2023.

    Several approaches have been introduced to understand surgical scenes through downstream tasks like captioning and surgical scene graph generation. However, most of them heavily rely on an independent object detector and region-based feature extractor. Encompassing computationally expensive detection and feature extraction models, these multi-stage methods suffer from slow inference speed, making them less suitable for real-time surgical applications. The performance of the downstream tasks also degrades from inheriting errors of the earlier modules of the pipeline. This work develops a detector-free gradient-based localized feature extraction approach that enables end-to-end model training for downstream surgical tasks such as report generation and tool-tissue interaction graph prediction. We eliminate the need for object detection or region proposal and feature extraction networks by extracting the features of interest from the discriminative regions in the feature map of the classification models. Here, the discriminative regions are localized using gradient-based localization techniques (e.g. Grad-CAM). We show that our proposed approaches enable the real-time deployment of end-to-end models for surgical downstream tasks. We extensively validate our approach on two surgical tasks: captioning and scene graph generation. The results prove that our gradient-based localized feature extraction methods effectively substitute the detector and feature extractor networks, allowing end-to-end model development with faster inference speed, essential for real-time surgical scene understanding tasks.

    Other authors
    See publication

  • Rethinking Feature Extraction: Gradient-Based Localized Feature Extraction for End-To-End Surgical Downstream Tasks

    IEEE Robotics and Automation Letters

    Several approaches have been introduced to understand surgical scenes through downstream tasks like captioning and surgical scene graph generation. However, most of them heavily rely on an independent object detector and region-based feature extractor. Encompassing computationally expensive detection and feature extraction models, these multi-stage methods suffer from slow inference speed, making them less suitable for real-time surgical applications. The performance of the downstream tasks…

    Several approaches have been introduced to understand surgical scenes through downstream tasks like captioning and surgical scene graph generation. However, most of them heavily rely on an independent object detector and region-based feature extractor. Encompassing computationally expensive detection and feature extraction models, these multi-stage methods suffer from slow inference speed, making them less suitable for real-time surgical applications. The performance of the downstream tasks also degrades from inheriting errors of the earlier modules of the pipeline. This work develops a detector-free gradient-based localized feature extraction approach that enables end-to-end model training for downstream surgical tasks such as report generation and tool-tissue interaction graph prediction. We eliminate the need for object detection or region proposal and feature extraction networks by extracting the features of interest from the discriminative regions in the feature map of the classification models. Here, the discriminative regions are localized using gradient-based localization techniques (e.g. Grad-CAM). We show that our proposed approaches enable the real-time deployment of end-to-end models for surgical downstream tasks. We extensively validate our approach on two surgical tasks: captioning and scene graph generation. The results prove that our gradient-based localized feature extraction methods effectively substitute the detector and feature extractor networks, allowing end-to-end model development with faster inference speed, essential for real-time surgical scene understanding tasks.

    Other authors
    See publication

  • Rich Feature Distillation with Feature Affinity Module for Efficient Image Dehazing

    Optik

    Single-image haze removal is a long-standing hurdle for computer vision applications. Several works have been focused on transferring advances from image classification , detection, and segmentation to the niche of image dehazing, primarily focusing on contrastive learning and knowledge distillation. However, these approaches prove computationally expensive, raising concern regarding their applicability to on-the-edge use-cases. This work introduces a simple, lightweight, and efficient…

    Single-image haze removal is a long-standing hurdle for computer vision applications. Several works have been focused on transferring advances from image classification , detection, and segmentation to the niche of image dehazing, primarily focusing on contrastive learning and knowledge distillation. However, these approaches prove computationally expensive, raising concern regarding their applicability to on-the-edge use-cases. This work introduces a simple, lightweight, and efficient framework for single-image haze removal, exploiting rich "dark-knowledge" information from a lightweight pre-trained super-resolution model via the notion of heterogeneous knowledge distillation. We designed a feature affinity module to maximize the flow of rich feature semantics from the super-resolution teacher to the student dehazing network. In order to evaluate the efficacy of our proposed framework, its performance as a plug-and-play setup to a baseline model is examined. Our experiments are carried out on the RESIDE-Standard dataset to demonstrate the robustness of our framework to the synthetic and real-world domains. The extensive qualitative and quantitative results provided establish the effectiveness of the framework, achieving gains of upto 15% (PSNR) while reducing the model size by∼20 times.

    Other authors
    See publication

  • Not All Lotteries Are Made Equal

    International Conference on Machine Learning (ICML)

    The Lottery Ticket Hypothesis (LTH) states that for a reasonably sized neural network, a sub-network within the same network yields no less performance than the dense counterpart when trained from the same initialization. This work investigates the relation between model size and the ease of finding these sparse sub-networks. We show through experiments that, surprisingly, under a finite budget, smaller models benefit more from Ticket Search (TS).

    Other authors
    See publication

  • Global-Reasoned Multi-Task Learning Model for Surgical Scene Understanding

    IEEE International Conference on Robotics and Automation (ICRA)

    Other authors
    See publication

  • CADSketchNet - An Annotated Sketch dataset for 3D CAD Model Retrieval with Deep Neural Networks

    Special Section on 3DOR2021 - 14th EG 3D Object Retrieval Workshop of the Computers & Graphics Journal

    The paper is accepted for publication in the Special Section on 3DOR2021 - 14th EG 3D Object Retrieval Workshop of the Computers & Graphics Journal.

    Other authors
    See publication

  • Audiomer: A Convolutional Transformer for Keyword Spotting

    AAAI Conference on Artificial Intelligence, 2022

    Other authors
    See publication

  • Global-Reasoned Multi-Task Learning Model for Surgical Scene Understanding

    IEEE Robotics and Automation Letters

    Other authors
    See publication

  • Introducing Self-Attention to Target Attentive Graph Neural Networks

    International Conference on Artificial Intelligence and Signal Processing '22

    Other authors
    See publication

  • User-Friendly Waveguide Modes Visualiser

    IEEE Microwave Magazine (IEEE MTT-S)

    Other authors
    See publication
Join now to see all publications

Projects

  • Fine-Tuning Pre-trained Language Models for Discrete Unit-based ASR

    -

    Our research combines self-supervised speech models (S3Ms) and large language models (LLMs) to enhance discrete speech unit-based Automatic Speech Recognition (ASR). By having S3Ms handle acoustic modeling and LLMs manage language modeling, we achieve superior results with up to a 21.7% relative word error rate reduction on LibriSpeech-100 compared to existing baselines.

  • Smoothed Gradient Descent-Ascent for Min-Max Optimization

    -

    - We leverage Smoothed Gradient Descent Algorithm (SGDA) to effectively tackle minimax optimization problems, mitigating convergence issues inherent in non-convex and non-concave scenarios.

    - SGDA outperforms Gradient Descent Ascent (GDA) across popular datasets, notably excelling in Generative Adversarial Networks (GANs) where it successfully avoids mode collapse and generates diverse, high-quality samples in MNIST, CIFAR100, and ImageNet200 experiments.

    - Explores SGDA's…

    - We leverage Smoothed Gradient Descent Algorithm (SGDA) to effectively tackle minimax optimization problems, mitigating convergence issues inherent in non-convex and non-concave scenarios.

    - SGDA outperforms Gradient Descent Ascent (GDA) across popular datasets, notably excelling in Generative Adversarial Networks (GANs) where it successfully avoids mode collapse and generates diverse, high-quality samples in MNIST, CIFAR100, and ImageNet200 experiments.

    - Explores SGDA's potential with more expressive models and dives deeper into a double-smoothed version of SGDA to improve generated sample diversity while pushing forward the realistic appearance of generated samples.

  • SurgIQ: Rethinking Visual Question Answering in Surgical Scenes

    -

    - SurgiQ revolutionizes Surgical Visual Question Answering (S-VQA) by leveraging Grad-CAM++ guided features for enhanced instance-level localization.

    - This approach eliminates the need for bounding box annotations and multi-stage processing, streamlining S-VQA while improving accuracy.

    - SurgiQ integrates Grad-CAM++ guided visual features, and Bio-BERT word embeddings (priors) with VisualBERT-ResMLP, resulting in robust visual-text representations for accurate answers.

    -…

    - SurgiQ revolutionizes Surgical Visual Question Answering (S-VQA) by leveraging Grad-CAM++ guided features for enhanced instance-level localization.

    - This approach eliminates the need for bounding box annotations and multi-stage processing, streamlining S-VQA while improving accuracy.

    - SurgiQ integrates Grad-CAM++ guided visual features, and Bio-BERT word embeddings (priors) with VisualBERT-ResMLP, resulting in robust visual-text representations for accurate answers.

    - The project simplifies the S-VQA pipeline with efficient inference and a straightforward classification head, offering a promising solution for medical domain challenges.

  • Compressing Vision Transformers for Low-Resource Visual Learning

    -

    - The project aims to optimize Vision Transformers (ViTs) for edge deployment on UAVs, targeting memory-efficient solutions without sacrificing accuracy.

    - Knowledge distillation from Swin-v2-T to MobileViT enhances IoU by a substantial 0.16, while ResNet18 + UPerNet improves by 0.12, showcasing the effectiveness of this technique.

    - On the Jetson Nano, MobileViT exhibits a 1.8x throughput increase when quantized to half-precision floats (fp16). However, ResNet18 maintains better…

    - The project aims to optimize Vision Transformers (ViTs) for edge deployment on UAVs, targeting memory-efficient solutions without sacrificing accuracy.

    - Knowledge distillation from Swin-v2-T to MobileViT enhances IoU by a substantial 0.16, while ResNet18 + UPerNet improves by 0.12, showcasing the effectiveness of this technique.

    - On the Jetson Nano, MobileViT exhibits a 1.8x throughput increase when quantized to half-precision floats (fp16). However, ResNet18 maintains better runtime due to PyTorch's convolution vs. attention kernel implementation.

    - The final framework (distilled and quantized UPerNet + MobileViT) consumes 3742 MB of RAM with 1030 MB of swap on the Jetson Nano, providing a promising framework for edge applications in surveillance, search-and-rescue, and more, with potential further optimization through bare-metal solutions and inference engines.

  • Attention-Guided Saliency-Enhanced Product Detector

    -

    - Designed an efficient, lightweight saliency detection framework (27.184 FPS, 7.45M params) for real-time retail item detection utilizing the attention mechanism for feature aggregation with edge refinement in the frequency domain.

    - Achieved a balance between performance and computational efficiency by utilizing masked edge, union, and object attention modules. This approach leads to a significant improvement in SOD performance, with a mean F-measure score of 0.875 on the DUTS dataset…

    - Designed an efficient, lightweight saliency detection framework (27.184 FPS, 7.45M params) for real-time retail item detection utilizing the attention mechanism for feature aggregation with edge refinement in the frequency domain.

    - Achieved a balance between performance and computational efficiency by utilizing masked edge, union, and object attention modules. This approach leads to a significant improvement in SOD performance, with a mean F-measure score of 0.875 on the DUTS dataset and significant improvements on in-house product datasets.

    - Designed an adaptive pixel intensity loss function that helps to improve the robustness of the network to noisy labels and improve local-global structure awareness.

  • Exploring the Lottery Ticket Hypothesis in Neural Networks

    -

    The Lottery Ticket Hypothesis states that neural networks contain much smaller sparse subnetworks capable of training to the same (or better) accuracy than the full network.

    - Applied Layer-Adaptive Magnitude Pruning (LAMP) to Lottery Ticket Hypothesis.

    - Showed that LAMP can be used to find Winning Tickets and analyze the distribution of weights.

    - Analyzed the expressive power of pruned networks with Trajectory Length and Activation Patterns v/s. their dense counterparts.

    See project

  • Algorithmic applications in 5G MIMO Systems

    -

    The contribution in this project is an experimental overview of designing algorithms from both software paradigms that are leveraged to provide solutions for various applications involving MMIMO technology and antenna design in 5G systems. Specifically:

    - A rule-based method for Obstacle Detection. This method involves MMIMO radar and the theory of cognition. (Software 1.0)

    - Learning algorithms for MIMO-based Antenna Applications - Machine Learning and Deep Learning algorithms…

    The contribution in this project is an experimental overview of designing algorithms from both software paradigms that are leveraged to provide solutions for various applications involving MMIMO technology and antenna design in 5G systems. Specifically:

    - A rule-based method for Obstacle Detection. This method involves MMIMO radar and the theory of cognition. (Software 1.0)

    - Learning algorithms for MIMO-based Antenna Applications - Machine Learning and Deep Learning algorithms used for Human Activity Classification, using human movement data captured by a radar device. (Software 2.0)

    𝐔𝐧𝐝𝐞𝐫 𝐑𝐞𝐯𝐢𝐞𝐰 𝐚𝐭 𝐈𝐄𝐄𝐄 𝐀𝐧𝐭𝐞𝐧𝐧𝐚𝐬 𝐚𝐧𝐝 𝐖𝐢𝐫𝐞𝐥𝐞𝐬𝐬 𝐏𝐫𝐨𝐩𝐚𝐠𝐚𝐭𝐢𝐨𝐧 𝐋𝐞𝐭𝐭𝐞𝐫𝐬

    Other creators

  • AI For Good - The "No Poverty Challenge"

    -

    - Analyzed extreme poverty population across the globe while comparing the efforts of the United Nations and other partners towards eradicating poverty.

    - Studied patterns in the data to draw inferences in an Unsupervised manner, building a predictive model to identify funding required by countries.

    𝐑𝐞𝐜𝐞𝐢𝐯𝐞𝐝 𝐚𝐧 𝐇𝐨𝐧𝐨𝐫𝐚𝐫𝐲 𝐦𝐞𝐧𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐭𝐡𝐞 𝐢𝐧𝐧𝐨𝐯𝐚𝐭𝐢𝐯𝐞 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧 𝐩𝐫𝐨𝐩𝐨𝐬𝐞𝐝, 𝐚𝐦𝐨𝐧𝐠𝐬𝐭 𝟏𝟎𝟎+ 𝐩𝐚𝐫𝐭𝐢𝐜𝐢𝐩𝐚𝐧𝐭𝐬.

  • Learning with Retrospection

    -

    The first PyTorch Implementation of Learning with Retrospection.

    - Learning with retrospection (LWR) makes use of the learned information in the past epochs to guide the subsequent training.

    - LWR is a simple yet effective training framework to improve accuracies, calibration, and robustness of DNNs without introducing any additional network parameters or inference cost, only with a negligible training overhead.

    𝐆𝐢𝐭𝐇𝐮𝐛 𝐂𝐨𝐝𝐞:…

    The first PyTorch Implementation of Learning with Retrospection.

    - Learning with retrospection (LWR) makes use of the learned information in the past epochs to guide the subsequent training.

    - LWR is a simple yet effective training framework to improve accuracies, calibration, and robustness of DNNs without introducing any additional network parameters or inference cost, only with a negligible training overhead.

    𝐆𝐢𝐭𝐇𝐮𝐛 𝐂𝐨𝐝𝐞: https://github.com/The-Learning-Machines/LearningWithRetrospection

    See project

  • MiraiShield

    -

    A steady increase in usage of IoT devices that are more vulnerable to security compromises than desktop computers has led to an increase in the occurrence of IoT-based botnet attacks. To mitigate this new threat, a deep learning model is designed in this project. This deep learning model uses a publicly available dataset for pre-training, which was further fine-tuned on scraped data and smaller datasets. Furthermore, an IoT network was designed, to attack the network with malware and show the…

    A steady increase in usage of IoT devices that are more vulnerable to security compromises than desktop computers has led to an increase in the occurrence of IoT-based botnet attacks. To mitigate this new threat, a deep learning model is designed in this project. This deep learning model uses a publicly available dataset for pre-training, which was further fine-tuned on scraped data and smaller datasets. Furthermore, an IoT network was designed, to attack the network with malware and show the effectiveness of the model.

    𝐆𝐢𝐭𝐇𝐮𝐛 𝐂𝐨𝐝𝐞 - https://github.com/smj007/MiraiShield
    𝐃𝐞𝐦𝐨 𝐕𝐢𝐝𝐞𝐨 - https://www.youtube.com/watch?v=In_BqB0dU_0

    See project

  • Generative Modelling - Architecture Zoo

    -

    Developed an experimental codebase for various papers researched, some of which include DCGAN, StyleGAN, InfoGAN, BiGAN, CycleGAN, DiscoGAN and LS-GAN.

    𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤𝐬: 𝐏𝐲𝐓𝐨𝐫𝐜𝐡 𝐚𝐧𝐝 𝐓𝐞𝐧𝐬𝐨𝐫𝐅𝐥𝐨𝐰

    See project

  • Waveguide Visualiser

    -

    A novel interactive Waveguide Visualizer using Python, depicting the complex Electric and Magnetic Vector fields in TE, TM, TEM modes, for Rectangular and Cylindrical waveguides. Presented to 100+ faculty of IITs and NITs in SPARC-2021. Published on Microwaves101. Was presented in the workshop titled - "Recent Trends on Metamaterial Antennas for Wireless Applications and Deep Learning Techniques", in association with the University of Saskatchewan, Canada.

    𝐀𝐜𝐜𝐞𝐩𝐭𝐞𝐝 𝐚𝐭 𝐭𝐡𝐞…

    A novel interactive Waveguide Visualizer using Python, depicting the complex Electric and Magnetic Vector fields in TE, TM, TEM modes, for Rectangular and Cylindrical waveguides. Presented to 100+ faculty of IITs and NITs in SPARC-2021. Published on Microwaves101. Was presented in the workshop titled - "Recent Trends on Metamaterial Antennas for Wireless Applications and Deep Learning Techniques", in association with the University of Saskatchewan, Canada.

    𝐀𝐜𝐜𝐞𝐩𝐭𝐞𝐝 𝐚𝐭 𝐭𝐡𝐞 𝐈𝐄𝐄𝐄 𝐌𝐢𝐜𝐫𝐨𝐰𝐚𝐯𝐞 𝐌𝐚𝐠𝐚𝐳𝐢𝐧𝐞 (𝐈𝐄𝐄𝐄 𝐌𝐓𝐓-𝐒)

    See project

  • Data-driven Low Altitude UAV Channel Model using Generative Adversarial Networks

    -

    - Demonstrated low altitude UAV Channel modeling using the data generated under four distinct environments, and hypothesize the use of a prior distribution of sample channel gains at different launch characteristics of a UAV.

    - Introduced the 𝑳𝑨-𝑼𝑨𝑽 𝑪𝒉𝒂𝒏𝒏𝒆𝒍 𝑫𝒂𝒕𝒂𝒔𝒆𝒕, consisting of normalized channel coefficients obtained from the four environments selected.

    - We verify the claim of modeling the channel using sample gains at a single angular level by developing a…

    - Demonstrated low altitude UAV Channel modeling using the data generated under four distinct environments, and hypothesize the use of a prior distribution of sample channel gains at different launch characteristics of a UAV.

    - Introduced the 𝑳𝑨-𝑼𝑨𝑽 𝑪𝒉𝒂𝒏𝒏𝒆𝒍 𝑫𝒂𝒕𝒂𝒔𝒆𝒕, consisting of normalized channel coefficients obtained from the four environments selected.

    - We verify the claim of modeling the channel using sample gains at a single angular level by developing a Vanilla GAN.

    - Further, we extend the system proposed for modeling the UAV channel to a cGAN, conditioned on launch angle information. We train the cGAN in different environments to ensure robustness to environmental change and any provided prior.

    𝐔𝐧𝐝𝐞𝐫 𝐑𝐞𝐯𝐢𝐞𝐰 𝐚𝐭 𝐭𝐡𝐞 𝐏𝐡𝐲𝐬𝐢𝐜𝐚𝐥 𝐂𝐨𝐦𝐦𝐮𝐧𝐢𝐜𝐚𝐭𝐢𝐨𝐧 𝐉𝐨𝐮𝐫𝐧𝐚𝐥

  • Deep RL agent to beat Atari's Breakout - Modified A3C

    -

    Implemented an Asynchronous Advantage actor-critic model (A3C) to ace the Breakout game by Atari, modified with an LSTM for improved experience replay.

    𝐆𝐢𝐭𝐇𝐮𝐛 𝐂𝐨𝐝𝐞: https://github.com/smj007/Breakout_A3C

    See project

  • Self Driving Car using Deep Q-Learning

    -

    Implemented an ANN to learn a path to be followed in a custom-drawn environment using the Kivy Language. Adopted the Q-Learning approach to solving the task at hand.

    𝐆𝐢𝐭𝐇𝐮𝐛 𝐂𝐨𝐝𝐞: https://github.com/smj007/Self-Driving-Car

    See project

  • Luminosity and Contrast Improvement in Color Fundus Images

    -

    Worked on Enhanced Fundus Imaging, to improve the Local Contrast and Luminance in Images of the Eye. Implemented the Discrete Wavelet Transform, and further implemented the Curvelet Transform as a part of the algorithmic approach to aid ophthalmologists, in easier detection of Exudates.

    𝐏𝐫𝐞𝐬𝐞𝐧𝐭𝐞𝐝 𝐚𝐭 𝐭𝐡𝐞 𝐒𝐢𝐠𝐧𝐚𝐥 𝐏𝐫𝐨𝐜𝐞𝐬𝐬𝐢𝐧𝐠 𝐋𝐚𝐛𝐨𝐫𝐚𝐭𝐨𝐫𝐲, 𝐍𝐈𝐓 𝐓𝐫𝐢𝐜𝐡𝐲

  • I-Travel: The one-stop travel ticket booking app

    -

    The project’s aim is to create a travel ticket booking application, which allows
    customers to book travel tickets with ease. Here the application 𝐈-𝐓𝐫𝐚𝐯𝐞𝐥 is depicted in such
    an attractive and user-friendly manner which makes the customer inwardly use the
    application.

    - By using Netbeans and its efficient feature JFrame as Form, the application is easily
    able to get sufficient information from the user and makes the process of ticket booking
    easier. Attractive…

    The project’s aim is to create a travel ticket booking application, which allows
    customers to book travel tickets with ease. Here the application 𝐈-𝐓𝐫𝐚𝐯𝐞𝐥 is depicted in such
    an attractive and user-friendly manner which makes the customer inwardly use the
    application.

    - By using Netbeans and its efficient feature JFrame as Form, the application is easily
    able to get sufficient information from the user and makes the process of ticket booking
    easier. Attractive backgrounds have also been attached that not only make the project look
    presentable but also enhance its functioning by luring more customers.

    - The Project is linked with Paytm so that for every referral and ticket booking made by the
    users, they gain some amount of money, which is credited to their Paytm account.

    - In order to make the application look more official, a unique website for 𝐈-𝐓𝐫𝐚𝐯𝐞𝐥
    has been used and the forms have been hyperlinked.

    - The usage of various components such as buttons, labels, text fields, radio
    buttons, combo boxes, date choosers, and text areas make the form easy to operate and the
    functioning is very transparent to the users.

    𝐏𝐫𝐞𝐬𝐞𝐧𝐭𝐞𝐝 𝐚𝐭 𝐭𝐡𝐞 𝐂𝐨𝐦𝐩𝐮𝐭𝐞𝐫 𝐒𝐜𝐢𝐞𝐧𝐜𝐞 𝐋𝐚𝐛𝐨𝐫𝐚𝐭𝐨𝐫𝐲, 𝐘𝐮𝐯𝐚𝐛𝐡𝐚𝐫𝐚𝐭𝐡𝐢 𝐏𝐮𝐛𝐥𝐢𝐜 𝐒𝐜𝐡𝐨𝐨𝐥

Honors & Awards

  • Reviewer

    IEEE Transactions on Neural Networks and Learning Systems

    Selected as a reviewer for IEEE Transactions on Neural Networks and Learning Systems (Impact Factor: 14.255), contributing to the advancement of knowledge in the field of applied ML.

  • Sri. Janardhana Iyengar Memorial Award

    National Institute of Technology, Tiruchirappalli

  • Institute Medal, 2022

    NIT Trichy

    Gold medalist, rank 1 in the Department of ECE

  • Research Week with Google

    Google Research India

    Shortlisted amongst 10000+ applicants to participate in Research Week with Google. Google Research India is hosting a four-day Research Week with Google. The program structure is to discuss the state of the art of a variety of ML Techniques and also discuss their limitations and critiques.

  • Dr. A.L Abdus Sattar Memorial Award

    National Institute of Technology, Tiruchirappalli

    Outstanding student in 3rd year, ECE

  • Institute Day Prize Winner (2021) - Rank 1 in ECE

    National Institute of Technology, Tiruchirappalli

    Awarded for 2nd year performance - Rank 1 in the Department of ECE

  • Mentorship Program

    Illuminate AI

    Selected for the 2021 cohort of the Mentorship Program at Illuminate AI. Working with peers to target 'AI For Good', developing Machine Learning pipelines to analyze poverty and hunger trends across the world, while deploying a mechanism to help deal with the crisis we are facing.

  • The Graphics Replicability Stamp Award

    Graphics Replicability Stamp Initiative

    Recognition of the service provided to the scientific community by releasing the CADSketchNet dataset for free, non-commercial use.

  • IASc-INSA-NASI Summer Research Fellowship

    Indian National Science Academy

    Nominated for the SRFP, Summer '21

  • DAAD-WISE Scholarship

    DAAD

    Selected for the WISE program in Germany, which targets Indian students pursuing a degree in Science/Engineering, and wish to do a research internship at a German Institute

  • AWS Machine Learning Scholarship

    AWS, Udacity

  • Annual Day Award for Academic Excellence

    Yuvabharathi Public School

    State Rank 1 in Informatics Practices (Computer Science), securing 100/100 in boards. Best outgoing student in the Engineering Stream of 2018

  • Japan Country Ambassador

    World Maths Day - 3P Learning

    Was selected as the representative for Japan, to help my peers and spread awareness, about the online competition, to ensure active and healthy participation.

  • World Rank 2

    World Maths Day - 3P Learning

    Finished as the category winner in Asia in a 48-hour long online international speed-math competition, with over 3,000,000 participants from 200+ countries

More activity by Sai Mitheran

View Sai Mitheran’s full profile

  • See who you know in common
  • Get introduced
  • Contact Sai Mitheran directly
Join to view full profile

Other similar profiles

Explore top content on LinkedIn

Find curated posts and insights for relevant topics all in one place.

View top content

Add new skills with these courses

See all courses