Processing, evaluating and tuning Deep Neural Networks
● ANN - Artificial Neural Networks
Dataset Description: A bank is investigating a very high rate of customer leaving the bank. Here is a 10,000 records dataset to investigate and predict which of the customers are more likely to leave the bank soon.
Classifier built as an experiment to learn about the different features of ANNs in Keras, with TensorFlow backend. The accuracies obtained have a lot of scope for improvement.
● CNN - Convolutional Neural Networks
Dataset Description: A clean, preprocessed image dataset containing 10000 cat and dog images. Training set of 8000 and test set of 2000 images.
Classifier built to experiment around with different features of CNNs in Keras. This helps in understanding how important it is to reduce the dimensions through convolutions and max pooling to better the accuracies.
Next up,
● RNN - Recurrent Neural Networks
Dataset Description: A Google stock price information for the years 2012 to 2016. It has 1258 readings for training. The test data is the stock price readings for the year 2017.
Regressor built to experiment the Keras LSTM model. The first regressor which has a single time step is improved to include 20 time-steps. This helps in understanding how a time-series regressor model is trained.
● SOM - Self Organized Maps
Dataset Description: A Credit Card Applications dataset from the UCI Machine Learning Repository. It has 690 instances of customer transactions, with 15 attribute features.
Classifier built to detect Customer fraud using minimalistic SOM model. This helps in understanding how SOMs work in an unsupervised manner and how they are different from Neural Networks with activation functions.
● Hybrid Model - SOM with ANN
Dataset Description: Same as that for SOM - Credit Card Applications
A hybrid Classifier built to achieve a model with better accuracies. Building a hybrid model helps in understanding the models individually (here, SOM and ANN) and how well the data can be processed to connect different models to obtain better ones.