Elvin Johnson Email & Phone Number

Pittsburgh, Pa, Us

Pittsburgh, Pennsylvania, United States

Creating machine learning/computer vision based products that transform warehouses. ML based predictive analytics for supply chain.

Pittsburgh, Pennsylvania, United States

Explainable Anomaly detection in Cytology/Cell Images (cervical cancer) using Machine Learning:- Built a novel end-to-end Deep Autoencoding Gaussian Mixture model (DAGMM) that would enable anomaly detection for images- Performed anomaly detection on cervical cancer pap smear test data with the DAGMM- Also incorporated anomaly detection for handcrafted features (features extracted from nucleus and cytoplasm) and performed anomaly detection with ECOD (Empirical Cumulative Distribution Functions) with the pyOD library- Combined anomaly predictions from the DAGMM and ECOD thus, introducing a multimodal pipeline for anomaly detection

Pittsburgh, Pennsylvania, United States

- Developed pipelines for processing, performing analysis/exploration of data lake data consisting of millions of rows, worth several TBs on Google Cloud Platform- Built ML models for segmentation, improving accuracy by 4%, created Bayesian models for estimating the uncertainty of prediction on CT/PET data for tumors - Combined clinical data with image data (radiomics) to train models like Random Forest, Linear Regression, KNN, SVR, Adaboost, Xgboost, to estimate patient recurrence free survival rate (RFS). Tuned models with grid search, evaluated them with various metrics- Built, trained, tuned ML classifiers with 100% accuracy - Invented a method for pose correction of reconstructed CBCT scans with ML, despite having scanty data, by generating synthetic data and training a UNet, achieving significant accuracy, robustness and effort cutback.

1. HeartSpot: Data Compression for Explainable Cardiomegaly Detection (Field: Machine Learning, Status: Accepted for publishing):Developed HeartSpot, a method for for Cardiomegaly detection based on domain knowledge which allows - 9x reduction in training time with a simple MLP, - privatization of X-ray data, - 32x compression of data, - achieves higher accuracy than a Densenet121 trained on uncompressed data. 2. DAGMM (Deep Autoencoding Gaussian Mixture models) for Anomaly Detection: - Built a DAGMM (Autoencoder followed by a Gaussian Mixture Model) in Pytorch that would enable End-to-End learning i.e gradients would backpropagate from a GMM through to a deep autoencoder. - The entire architecture developed facilitates an unsupervised learning method for the detection of anomalies in a dataset. - These anomalies/outliers/bad examples could them be removed from the dataset, allowing models to train faster, more efficiently and achieve higher accuracy. 3. DeepFixCX: Explainable Privacy-Preserving Image Compression for Medical Image Analysis (Field: Machine Learning, Status: Accepted for publishing):DeepFixCX is a ML based method for image compression, where the compression can be controlled using explainable hyperparameters, ensures privatization of data, and also increases the training speed by over 70 times, without reduction in accuracy. My main tasks in this project were to handle - model creation and training, - hyperparamter tuning, - model regularization - model evaluation.

Signature Extraction from Documents using ML1. Developed an end-to-end solution to automate this task for a client and managed the entire life cycle of the data science project 2. Preprocessed, collected, cleaned, augmented data, trained various ML models like Fast RCNN Inception, SSD Resnet50 etc to identify, extract signatures, performed model evaluation using loss, mAP, designed a local web app/UI for deploying the trained ML model3. Reduced time for handling signatures by 90% via the final app, which allows for uploading a document and returns extracted signaturesTruck Load Optimization for Logistics1. Executed truck load optimization (logistics) and automation for a client; performed data assessment, cleaning, merging and wrote code to generate useful reports from this data, thus, enabling optimized/automated truck loading, achieving 96% reduction in overall effort2. Extracted features from vendor data and conducted vendor performance assessment for logistics with the team, improving efficiencyHR Analytics

Mumbai Metropolitan Region

Image Analytics in Retail using ML1. Collaborated with the data science team to identify relevant use cases for image analytics in retail; 2. Initiated projects on building solutions, collected data, trained ML models with an accuracy of 87% 3. documented and published a white paper on ‘Image Analytics in Retail’ here

Mumbai, Maharashtra, India

1. Processed and cleaned about 1 million frames from deep fake videos (500 GB) by Facebook, Faceforensics, Google for building a dataset2. Developed a method to detect deep fakes via Transfer learning, CNNs and attained an accuracy of 90% on the test set 3. Conducted training, processing of ML models on the Nvidia DGX-1 AI Supercomputer and finished amongst the top 15% in Kaggle’s ‘Deep Fake detection Challenge’ by Facebook4. Presented and published findings as ‘Suratkar Shraddha, Elvin Johnson, Karan Variyambat, Mihir Panchal, and Faruk Kazi. "Employing Transfer-Learning based CNN architectures to Enhance the Generalizability of Deep fake Detection." In 2020 11th International Conference on Computing, Communication and Networking Technologies (ICCCNT), pp. 1-9. IEEE, 2020’, held at IIT Kharagpur, July 2020.

Mumbai Area, India

Successfully constructed a deep learning model based on Medical Imaging that would help in skin cancer detection. It would provide a two-fold utility whereby it would :1. Identify whether the tumor is benign or malignant.2. Classification of tumor into seven distinct categories and predictingthe most probable out of those seven.Several different deep learning models were created and all them were tested for accuracy.

Master'S Degree, Computer Engineering (Focus On Machine Learning And Data Science), 3.93/4.0

Btech - Bachelor Of Technology, Computer Engineering, 9.04/10 (Cgpa)