PruthviRaj G
Graduate Student at CMU | Ex - Technical Leader
Deep Learning, Machine Learning, Autonomous Driving, and Computer vision Enthusiast
Experienced Professional with 4+ years with a demonstrated history of working where applying Machine Learning to the Automotive Industry has been my forte. Skilled in M-scripting, Python, Pytorch, Tensorflow, MATLAB, and SIMULINK, C/C++, Model-Based Design, Embedded C, and Simulink, Excel VBA. I have worked on many exciting problems throughout this journey, such as Visual Obstruction Detection (CV), Customer Query Rerouting System (NLP), Pressure Adaptable Braking System, and many more. After coming to CMU, I narrowed down my focus on Deep Learning, specifically Computer Vision and Autonomous Driving, where I worked on interesting projects like Lane Recognition with Instance Segmentation, Face Classification and Verification, Weakly Supervised Audio Event Detection, End-to-end Frame level Speech Phoneme Classification, and Emotion Recognition using Spatio-Temporal data. A strong engineering & business development professional with an under-graduation Bachelor of Technology (B.Tech) in Electronics and Communications Engineering from SRM University - Merit Scholarship (50% Fee wavier) - ALL India SRMJEE 98th Rank Holder. Pursuing master's in Carnegie Mellon University (CMU) in ECE with a concentration in AI/ML Systems, focusing on Deep Learning, Machine Learning, Autonomous Driving and Computer vision.
personal info
Fall 2021 : 18786 - Intro to Deep Learning, 18794 - Patter Recognition Theory, 18793 - Image and Video Processing. Spring 2022: 11777 - Multimodal Machine Learning 16720 - Computer Vision 24678 - Computer Vision for Engineers Research Assistant - Cylab - Object Detection and Segmentation (Under Prof. Marios Savvides) Teaching Assistant - Computational Techniques (Under Prof. Elias Towe)
- Rear Camera Obstruction Detection: Implemented and delivered a system using CNNs to detect any obstructions in the rear-view camera to notify the driver for potential crashes while reversing a vehicle.
- Customer Query Classification and Rerouting: Trained and deployed an NLP solution to understand and reroute a customer’s natural language queries to respective departments. This solution helped in reducing the front-end workforce by 70% in a span of 3 months.
- Automatic Safe Breaking System: Developed a CNN model to detect obstacles using a front camera that helps in distinguishing soft passable objects from hard impassable objects. Helped in mitigating the vehicle’s frontal damage by 60% by applying breaks for obstacles.
- Pressure Adaptable Braking System: Formulated a deep neural network-based mechanism to adapt the pressure applied on breaks depending on the surface the vehicle is traveling. This system reduced the brake pressure mismatch accidents by 80%.
- Automated Dataset creation for BMW: Spearheaded a team for collecting and annotating 80K images for downstream applications like segmentation and classification. This data helped in increasing the mean average precision (mAP) by 40%.
Technical Leader (Automotive Tools Developer) with 4 years of experience in MACHINE LEARNING and DEEP LEARNING.My work involves using PYTHON, MATLAB, KERAS along with different machine learning algorithms depending on the problem at hand
- LANE RECOGNITION FOR AUTONOMOUS DRIVING USING DEEP NEURAL NETWORKS
- Devised and built a shared encoder, end-to-end encoder-decoder model for simultaneous lane detection and identification. Pixel wise binary classification.
- Cross Entropy Loss is used to train the Lane detection submodel, and Lane Identification subnet is trained using discriminative loss to optimize for better cluster separation.
- Trained and evaluated the above architecture on TuSimple dataset using accuracy, recall, precision, and F-score. Attained an accuracy of 96.35% and 43.2% F-score for pixel-wise detection and identification.
- WEAKLY SUPERVISED AUDIO EVENT DETECTION USING N-VIEW MULTITASK LEARNING
- Designed label noise robust Neural-Net architecture to learn Audio events from weakly labeled web data. Analyzed the effects of label noises (uniform, class and feature dependent noise) on the model performance.
- Implemented a 2-view architecture and extended to n-view multitask-learning model with nth view of input generated using CNN trained on full Audio Set data. Achieved classification accuracy at 43.2 MAP with subsets of Audio Set data (40 classes) as input.
- Visualized the data evaluation metrics using wandb framework by training multiple experimental setups with and without noise.
- FACE CLASSIFICATION AND VERIFICATION USING PROXY BOTTLENECK CNN MODEL
- Implemented and trained a ResNet-34 inspired Proxy bottleneck CNN model to classify a given face image into 4000 classes while preserving crucial discriminative features. Accomplished a test accuracy of 87.9%.
- Repurposed the above face classification model to perform face verification task by utilizing the embeddings from the bottleneck layer. Cosine similarity is measured between two face’s embeddings for verification. Obtained an AUC-ROC of 94.79%.
- END-TO-END FRAME-LEVEL SPEECH PHONEME CLASSIFICATION
- Developed a deep feed-forward neural network to classify every 10 ms MFCC (Mel-Frequency Cepstral Coefficients) window to 40 phoneme classes. Explored context-based modeling with horizontally stacked frame-level contextual input.
- Achieved a test accuracy of 80.92%.
- MULTIMODAL EMOTION RECOGNITION WITH SPATIO-TEMPORAL CLASSIFICATION
- Performed a comparative analysis for emotion classification using audio modality with CNNs, bidirectional (RNNs, GRUs and LSTMs) with highest accuracy of 75.3% using bidirectional GRU’s on RAVDESS dataset.
- Achieved an accuracy of 86% on video modality with horizontally stacked images as input to CNN’s which preserves the temporal relationship (change of emotion of actor in the video with time).
- Built a multimodal late fusion model using bidirectional GRU’s for audio, CNN for video and MLP for fused vector.
- LICENSE PLATE RECOGNITION USING OCR
- Created and engineered a real-time license plate identification and extraction pipeline that would acquire the image, preprocess into required format, segment the area of interest, and execute optical character recognition to acquire the license plate number for a toll collection center.
- Attained a character error rate of 2% at 30 frames per second
“Realization of Portable Data Acquisition System for Rocket Motor Static Test.” (PruthviRaj Gampalwar, P. Sandeep, K. Nikhil Reddy, and Dr.J. Manjula, 2017, Aug). In International Conference paper on Automation, Robotics and Mechatronics – ICARM 2017(pp. 89) – Conference paper presentation.