MACHINE LEARNING RESEARCH

Fedo

SCOPE OF PROJECT

Neural Network Analysis and EDA

Timeline: January - March 2021

What We Did:  Data Augmentation, Model Exploration, Model Recommendation

Tools: Python, Sckit-Learn

Goal of Project

The UCLA Data Science Union looks to perform exploratory data analysis on any data Fedo has on the impact of its product for its clients. Afterwards, DSU looks to improve any aspects of Fedo’s neural network that could use fine-tuning, troubleshooting, or could otherwise be improved upon.

What is Fedo?

Fedo is a technology company based in Bangalore, India that has created a proprietary machine learning algorithm to output a “Fedo score” on a scale of 0-1000 based on the input of an image of a client’s face. The score provides a metric for insurance companies to more accurately underwrite policies for their clients. 

Phase 1: Explore Fedo’s baseline model and analyze it’s performance

Phase 2: Data Augmentation (Face Cropping using Recognition Library, Applying Hessian Filter, Applying Laplacian Filter)

Cropping Images: The Method

  • Use HOG (histogram of oriented gradients) to detect face borders

  • Expanded borders by 25% to include surrondings (ex. hair, beard, ears)

  • Cropped images down to faces and surroundings, resized to 256x256 square

Hessian Filter: The Method

  • Used Hessian Filter from scikit-image to detect ridges

  • Solved for eigenvalues of Hessian matrix to detect continuity

Laplacian of Gaussian: The Method

  • 2nd derivative works as edge detector

  • Used Gaussian blur to reduce noise caused by 2nd derivative’s sensitivity

Phase 3: Analyze Performance Metrics

Their baseline model was performing at a validation accuracy of 52.78% on our given dataset. We decided to explore various models and tune the parameters to recommend the best model for Fedo

Phase 4: Recommend Best Model for Fedo

We decided to use the models VGG16, VGG19, ResNet, Inception v3. We followed the same data augmentation process, and our best performing model was Inception v3, with a Training Accuracy of 83.14% achieved in combination with cropping and fine tuning. 

Our Impact

We have improved Fedo’s modeling abilities to categorize the health score for potential insurance clients. As the flagship portion of their business, they have continued to expand and build upon offering their services to insurance companies with the help of DSU.