Iype Eldho - Projects

Machine Learning Researcher at Active Robotics Sensing Lab, NCSU (2023-24)

Developed a deep learning model to classify audio as speech or cough and detect Out-of-distribution samples using transformer-to-CNN knowledge distillation, improving baseline CNN model accuracy from 91% to 98%.
Utilized Decoupling MaxLogit and Virtual Logit Matching OOD detection methods with neural networks like WideResNet and DenseNet, achieving DenseNet AUROCs of 90.85% for OOD-Human and 82.34% for OOD-Other categories, with ViM showing superior performance.
Thesis: "Audio-Based Detection of Speech and Cough Using Out-of-Distribution Techniques for Filtering Similar and Different Kinds of Sounds."
Developed a multimodal cough and speech detection and enhanced it using an Out-of-Distribution algorithm maintaining 92% accuracy after adding the OOD samples.
Publication: “Robust Multimodal Cough and Speech Detection using Wearables: A Preliminary Analysis” (Publication Link)

Analysis of Unsupervised Domain Adaptation Models (2022)

Modelled Transformers and Neural Networks to assess image classification performance across domains, yielding a top F1 score of 97%, from the baseline of 76%.

Github Link | Project Poster

Geographic Terrain Identification for Prosthetic Limb (2022)

Produced a PyTorch-based LSTM model that utilized accelerometer and gyroscope data from a prosthetic limb to identify geographic terrain types, achieving the top F1 score of 88% from a baseline of 67%.

Github Link

Cross-lingual Emotion Recognition from Speech (2022)

Devised a Domain Adversarial Neural Network to identify emotions from a Mandarin Speech Dataset using an English dataset, achieving F1 scores of 97% & 60% in supervised and unsupervised models (20% boost).

Project Poster

Heartbeat Anomaly Detection (2022)

Generated a TensorFlow-based pipeline to detect arrhythmia in heartbeat sounds using SVM and LSTM models, with 78% F1-Score. Optimized the pipeline through cross-validation and Grid Search with 84% accuracy.

Github Link | Project Poster