spratyush02

toxic_comment_classification

Age-Gender-Ethnicity-Classifier

In this project we aim to explore a dataset of images of people from different backgrounds and build a neural network image classifier to classify the age, gender and ethnicity of people in images. We deal with three different tasks in the project. These are gender classification which is a binary classification problem while age prediction is a regression problem since the output of the network can be any discrete number. Our last problem is that of ethnicity classification which is a multi class classification problem. All these tasks are important and useful in several fields and hence developing an efficient classification model can be helpful in solving the problems. For example, classifying ethnicity is an important task to solve for federal authorities and many companies developing photo editing applications.

docsum-chat

Human-Motion-Prediction-RNN-SPL-

This project was done as a part of Machine Perception coursework in spring'20 at ETH Zurich and resulted in top of the leaderboard.

mouse_mover

Mouse cursor mover

Reliable-and-Interpretable-AI-Robustness-against-Adversarial-Attacks-

This project was done as a part of coursework "Reliable and Interpretable Artificial Intelligence" at ETH Zurich

RSNA-Intracranial-Hemorrhage-Detection-using-Deep-Learning

We present a method to correctly predict presence of Intracranial Hemorrhage and identify its type. Our model imitates the procedure followed by radiologists to analyse a 3D CT scan in real-world. The model utilizes multi-window 3D context from neighboring slices to improve predictions at each slice and subsequently, aggregates the slicelevel predictions to provide patient diagnosis for Intracranial Hemorrhage . Our proposed architecture performs significantly better than standard single window based non-contextual models

spratyush02

Config files for my GitHub profile.

time-series-forecasting

In this project, we analysed the timeseries for the forecasting of Coal prices in South Africa. We did extensive 'Exploratory Data Analysis' with the help of which we performed data cleaning, processing along with feature selection and extraction. We checked the timeseries for stationarity and seasonality using both graphical and mathematical models. We handled outliers in some years. We believe the reason for the outliers is the economic crisis in 2008. There was a sharp slowdown in demand, and with mining output remaining stubbornly high. Therefore, coal benchmarks fell down. We trained Classical timeseries forecasting method: ARIMA (Auto Regressive Integrated Moving Average), and a Deep Learning method using LSTM along with extensive hyperparameter search using TALOS library. We calculate the metric R^2 (coefficient of determination) regression score function. The R^2 score obtained using ARIMA is 0.96 in comparison to 0.61 which is achieved by LSTM model. For this, not very complex, timeseries ARIMA outperforms the LSTM model as we are trying an overly complex Deep Learning method to fit on a simple timeseries data.

toxic_comment_classification

In this work, we present three learning models, namely, a simple linear baseline based on logistic regression, a bidirectional gated recurrent unit based convolutional network with GloVe embeddings, and a BERT model for toxic comment classification

spratyush02/Human-Motion-Prediction-RNN-SPL-

This project was done as a part of Machine Perception coursework in spring'20 at ETH Zurich and resulted in top of the leaderboard.

spratyush02/RSNA-Intracranial-Hemorrhage-Detection-using-Deep-Learning

We present a method to correctly predict presence of Intracranial Hemorrhage and identify its type. Our model imitates the procedure followed by radiologists to analyse a 3D CT scan in real-world. The model utilizes multi-window 3D context from neighboring slices to improve predictions at each slice and subsequently, aggregates the slicelevel predictions to provide patient diagnosis for Intracranial Hemorrhage . Our proposed architecture performs significantly better than standard single window based non-contextual models

spratyush02/time-series-forecasting

In this project, we analysed the timeseries for the forecasting of Coal prices in South Africa. We did extensive 'Exploratory Data Analysis' with the help of which we performed data cleaning, processing along with feature selection and extraction. We checked the timeseries for stationarity and seasonality using both graphical and mathematical models. We handled outliers in some years. We believe the reason for the outliers is the economic crisis in 2008. There was a sharp slowdown in demand, and with mining output remaining stubbornly high. Therefore, coal benchmarks fell down. We trained Classical timeseries forecasting method: ARIMA (Auto Regressive Integrated Moving Average), and a Deep Learning method using LSTM along with extensive hyperparameter search using TALOS library. We calculate the metric R^2 (coefficient of determination) regression score function. The R^2 score obtained using ARIMA is 0.96 in comparison to 0.61 which is achieved by LSTM model. For this, not very complex, timeseries ARIMA outperforms the LSTM model as we are trying an overly complex Deep Learning method to fit on a simple timeseries data.

spratyush02/Age-Gender-Ethnicity-Classifier

In this project we aim to explore a dataset of images of people from different backgrounds and build a neural network image classifier to classify the age, gender and ethnicity of people in images. We deal with three different tasks in the project. These are gender classification which is a binary classification problem while age prediction is a regression problem since the output of the network can be any discrete number. Our last problem is that of ethnicity classification which is a multi class classification problem. All these tasks are important and useful in several fields and hence developing an efficient classification model can be helpful in solving the problems. For example, classifying ethnicity is an important task to solve for federal authorities and many companies developing photo editing applications.

spratyush02/docsum-chat

spratyush02/mouse_mover

Mouse cursor mover

spratyush02/Reliable-and-Interpretable-AI-Robustness-against-Adversarial-Attacks-

This project was done as a part of coursework "Reliable and Interpretable Artificial Intelligence" at ETH Zurich

spratyush02/spratyush02

Config files for my GitHub profile.

spratyush02/toxic_comment_classification

In this work, we present three learning models, namely, a simple linear baseline based on logistic regression, a bidirectional gated recurrent unit based convolutional network with GloVe embeddings, and a BERT model for toxic comment classification