Developed an AWS Rekognition model employing machine learning models to detect image labels, text, and recognize celebrities from images on the web and classified image context from the labels, text, and people by setting thresholds for different confidence levels.
In this project, the performance and efficiency of an authentication system based on keystroke dynamics was improved by comparing different feature extraction techniques and various machine learning techniques. Inferences were drawn to understand why some techniques worked well and some failed. Experiments were based on understanding the data in terms of being able to determine which techniques might prove to be effective.
Developed a deep learning model (less than 15 layers) with comparable validation accuracy of Xception model (more than 100 layers) with no statistical significance difference between classifier accuracies. This model was trained on ship dataset with five classes of ships and the data was imbalanced with more than 70% of total ship examples belonging to 'cargo' class. Provisions were made to deal with this imbalance without degrading the model performance and generalization accuracy.
Developed a Face Recognition system using dimensionality reduction and machine learning and compared different machine learning models to find the model that reduces the generalization loss the most. For this model, different features of the face were analyzed to identify the most discriminating features leading to high validation accuracy.
Fine-tuned the existing model to improve the accuracy of the old model from 76% to 89.4% and analyzed different parameters and hyperparameters of a convolutional neural network. The secondary aim of the project was to familiarize with different hyper parameters of a deep learning model and to facilitate understanding of how these parameters affect model performance.