Human-Activity-Recognition-using-LSTMs-Opportunity-UCI-Dataset-

Human Activity Recognition using LSTMs - Deep Learning

Successfully designed and trained several neural network models using LSTMs (Long Short Term Memory networks) for recognizing Human Activities by taking raw data from the sensors as a part of the course Neural Networks and Fuzzy Logic.

1) Opportunity dataset was used, which required a high amount of preprocessing like differentiating data into training, cross-validating and testing data, handling clusters of NAN values, converting data to its absolute values, eliminating certain similar columns based on their correlation values and converting 2-D data to 3-D for easy training of LSTMs.

2) Hyper-parameter tuning was accomplished by choosing proper dropout, window size for LSTM input, no. of layers and nodes per layer, optimizing function, early stopping, no. of epochs, etc.

3) Models were cross-validated and tested. Various evaluating parameters were calculated which gave accuracy in the range of 80% to 95% achieved an accuracy of 94.8% on the best-trained model.