This section includes all projects done on online platforms like Coursera.

Supervised Learning using Deep Neural Networks

Coursera Course: Neural Networks and Deep Learning Certification
Offered by: deeplearning.ai in their Deep Learning Specialization

• Implemented a single hidden layer shallow neural network for binary classification of different types of data sets
• Trained the model for a dataset comprising of dots in the shape of a flower to get an accuracy of 91.25%
• Implemented a deep neural network for binary classification of Cat vs Non-Cat images with an accuracy of 80%
• Implemented ‘He initialisation’ on a 3 layered DNN for binary classification of red and blue dots with an accuracy of 99%

Art Generation with Neural Style Transfer

Coursera Course: Convolutional Neural Networks Certification
Offered by: deeplearning.ai in their Deep Learning Specialization

• Used Neural Style Transfer on top of a pretrained VGG-19 network
• Merged style cost from multiple hidden layers and tuned the weights to get a final style cost function
• Defined total cost as a linear combination of content cost and style cost
• Used Adam optimizer to optimize the cost function to get the pixel values of the generated image
• The generated image combined the content of an image with the style of a painting
• Mixed image of the Louvre museum in Paris (content image C), with a painting by Claude Monet, a leader of the impressionist movement (style image S)

Car Detection with YOLO

Coursera Course: Convolutional Neural Networks Certification
Offered by: deeplearning.ai in their Deep Learning Specialization o Dfdfdff o ddvv

Face Recognition and Face Verification

Coursera Course: Convolutional Neural Networks Certification
Offered by: deeplearning.ai in their Deep Learning Specialization

• Used Keras to build a facial verification model for identifying whether the employee is having the correct ID Card and is indeed an employee
• Built upon a pretrained FaceNet model
• Implemented a triplet loss function to distinguish between positive and negative from the anchor
• Built a facial recognition model which eliminated the need of ID Card for the employees
• The facial recognition model ran for all the images in the database and reported the database encoding which had the least L2 distance from the image encoding
• The facial verification was a 1:1 problem whereas the facial recognition was a harder 1:K problem