Mitigating bias in Digital Media Content by Classifying Toxic and Profane Comments

• • Developed classical machine learning and ensemble models to identify the usage and degree of toxicity in the online content and to mitigate the bias of the models towards such toxic content.
• • Used the Wikipedia data set to Train and test the model.
• • Used Logistic regression as the baseline model and compared its performance with the Multinomial Naive Bayes,Random Forest and a proposed Ensemble model which consisted of pretrained logisitc regression model and adaboost model.
• • Performed multi-class classification and classified the sentences into six classes as per the degree of the toxicity.
• • Built the feature space using TF-IDF and count vectoriser.
• • Used libraries such as spacy,gensim sklearn, nltk,textblob, scrapy, keras to implement the models.
• • The proposed Ensemble model outperformed all the other models and was able to classify with an accuracy of 95.3
• • Random Forest model was the second best performing model among all the models.

Text Summarization

• • Worked in a team of three in NLP course to develop models for the Text Summarization models. Developed thebaseline model using the Extractive Text Summarization technique and compared its performance with the models based on Abstractive text summarization approach.
• • Implemented the baseline model by ranking sentences in the order of their importance and forming a summary with most relevant sentences.
• • Implemented a Transformer model and Seq2Seq model, consisting of a Bidirectional LSTM encoder and LSTM decoder with the local attention for the Abstractive approach.
• • Calculated the ROGUE scores of all the three models to evaluate the performance of the models. Transformers outperformed both the other models.

Ping Pong Game

• • Implemented a pingpong game based on deep reinforcement learning and implemented Convolutional Neural Network for the visual and score analysis.Created an agent environment loop where the agent gets better and better at the game by Trials and Errors as the model is trained overtime.
• • Used Deep Q learning which updates it Policy (action state mapping) based on the reward.
• • Used Convolutional Neural Network to analyse the inputs which were the window frame pixels and the score.
• • Used pygame library to design the entire game and keep the track of the scare tally.
Reinforcement Learning

Unsupervised Image Segmentation

• • Worked in a team of three in Computer Vision course to implement Unsupervised image segmentation using back propagation.
• • Improvised a joint learning approach that predicts, for a random image input, undetermined cluster labels and learns optimal CNN parameters for the image pixel clustering.
• • Implemented the network with 4 convolution layers with alternating 3x3 and 1x1 kernels and did batch normalisation.Used ReLU as the activation function.
• • Implemented the Felzenszwalb Algorithm for the pre-segmentation of the images.
• • The model was more accurate than the original proposed network model and was also faster by a factor of 10.
Computer Vision

Image Captioning

Trained the model using deep learning methods with several hundred of hyperparameters.

Interactive Data Visualization

Data Visualization is done using R and many of its libraries such as ggplot2, plotly, shiny, etc.

Prediciting the innovation rate in a disaster struck area

The project is developed in R language. It predicts the innovation rate in a disaster struck area.