Problem - This is a binary classification problem where observation of environmental factors such as temperature and humidity is used to classify whether a room is occupied or not.
Dataset - https://archive.ics.uci.edu/ml/datasets/Occupancy+Detection+
Result - Recurrent neural network (LSTM) model to predict occupancy of a room, 98% accuracy on the test set compared to a baseline accuracy of 94% for a softmax classifier
Problem - This is a multiclass classification problem where an image of a plant seedling has to be classified as one of the 12 species.
Dataset - https://www.kaggle.com/c/plant-seedlings-classification/data
Result - A simple CNN with accuracy around 70%
3. Seedling Classification with model tuning on Amazon SageMaker
Problem - This is a multiclass classification problem where an image of a plant seedling has to be classified as one of the 12 species.
Dataset - https://www.kaggle.com/c/plant-seedlings-classification/data
Result - Performed transfer learning (VGG19 model) and hyperparameter optimization using AWS Sagemaker (build your own container for Keras) and got an accuracy of 83.76%
4. Music Mood Classification using Fast.ai and Pytorch [Work in Progress]
Building a music mood classifier that classifies a music clip to one of 6 moods (happy, sad, angry, scary, funny, tender) by converting audio to image (Spectrogram, Melspectrogram, Tempogram) and using fastai library, transfer learning, pytorch to train the mode
Problem - This is a multiclass classification problem where an audio is classified based on its mood, as one of 6 classes (happy, sad, angry, scary, funny, tender).
Dataset - https://research.google.com/audioset/dataset/index.html
Result - Performed transfer learning (Resnet50) on Melspectrogram using fastai library and got an accuracy of 66.84%. Removed the ambiguous mood/class - 'Tender' and got an accuracy of 79%