Vijaya Krishna

The objective was to recognize thirty short word utterances recorded from people differing in sex, age and nationality.

1. The frequency data extracted from the audio signal were converted to Mel-frequency cepstral coefficients (20 x 11 x 1 matrix). As all the audio files were exactly one second long, padding of the data matrix was not neccessary.

2. Tried Convolutional neural network, Long short term memory network and Gated recurrent units network inputting the extracted features. As there was no imbalance in the data, accuracy score was itself a good evaluation metric. Achieved the highest accuracy score of 0.913 on the test data using GRUs.

In this project,the selected business questions were : Who were the revenue generating customers? ; How much revenue are they generating? ; What is their lifetime value to the business?

1. To identify the revenue generating customers, a combination of Logistic regression, random forest and gradient boosting classifiers were implemented resulting in a recall score of 0.885

2. A combination of Linear regression, random forest and gradient boosting regressors were implemented to forecast the revenue generated through the identified customers resulting in a RMSLE score of 1.625

3. For the monetary based ranking analysis, the Betageofitter and GammaFitter models were implemented

The objective of this project was to build a system to generate topics from the given corpus and further generate meaningful headlines depending on the input from the user.

1. Latent Dirichlet Allocation (LDA), Non-Negative Matrix Factorization (NMF) and Markov Chain models were implemented to generate topics

2. Long Short Term Memory (LSTM) networks was implemented to generate meaningful headlines depending on the input from the user

The goal of the project was to build a system that identifies toxic comments in social media platforms and categorize them into one or mor of the classes: toxic, severe toxic, obscene, threat, insult and identity hate.

1. As part of data pre-processing, stopwords, punctuations, blank lines, URLs, hyperlinks were removed. Used NLTK's WordNetLemmatizer to lemmatize the words and used Glove 100d word vectors as embeddings.

2. Used Tensorflow and Sci-kit learn to build a Convolutional Neural Network, Recurrent neural network and Naive bayes SVM. The model outputs were stacked to get the class predictions.

The requirement was to project the sales of both newly launched and existing Denso robots in India by providing a month ahead and real time daily updates of the forecasts.

1. For the newly launched items, 12-month forecasts were provided using an ensemble of L2-Regularized regression and a fixed effects model by considering the trend and seasonality observed in robots with similar characteristics such as cost, payload, degree of freedom and IP code.

2. For the existing items, after identifying new interactive features and basic transformations, a 3-month forecast with real time daily updates was provided using a L2-Regularized regression.

3. Tableau was used to represent and deliver the results

The goal of the project was to build a system that detects defective casting with rifes and holes using image data captured from a vision system.

1. Used Gaussian mixture model for clustering.

2. Extracted features using pre-trained VGG16 after removing the output layer. Used data augmentation to counter the imbalance in the data. Used Convolutional neural network to classify the defective castings from the good ones.

3. Deployed the system in the robotic controller and monitored the system in production for accuracy for one month.

The goal of the project was to come up with a function that predicts the market value of a player when their attributes are given as inputs. Helps team managers to make better and informed player transfers and negotiate wage based on a certain player's overall abilities.

1. Performed exploratory data analysis to better understand the data. Manually removed obvious features that will contribute nothing to the model like team flag, club logo, name etc. Manually categorized ordinal and nominal variables and encoded them accordingly.
2. Used RandomizedLasso as the stability selection method to select the top 30 strong features.
3. Used ridge regression, random forests and SVR to predict the market value. The lowest mean square error of 1.86 was achieved after fine tuning the random forests model.

The goal of the project was to build a model that learns a mapping from the videos and captions in a semantic space such that we can compare the semantic content by comparing the embeddings.

1. There are two parts in the feature extraction process. Extracted frame embeddings by passing the frames in the ActivityNet videos through a pre-trained AlexNet and then applied Principal Component Analysis to further reduce the feature size to a 500 dimension vector for each frame. To extract the caption embeddings, the captions sourced from Stanford's Vision lab research project corresponding to ActivityNet videos were passed through Facebook's FastText model to get a 300 dimension vector for each caption.

2. The model used was a modified version of a Siamese network with two recurrent network architectures. The two sets of extracted embeddings were passed to their corresponding network architectures that donot share their parameters. Trained the network using triplet and pair wise loss function.