Implementation of a fully connected feed-forward neural network using the back propagation algorithm for stochastic batch gradient descent computations. Network size is adoptive and supports MLP.
Activations: softmax, sigmoid
Loss functions: log-likelihood, mean-square, cross-entropy(binary equivalent of log-likelihood)
Regularization: L2
Validation - Takes labels and data as input
Hyper-parameters: learning-rate(eta), regularization-parameter(lambda), epochs
Used for training and deploying human activity recognition system using wearable sensors.
Accuracy on test set using 6-9-6 neural network for multi-class classification using softmax layer: 97.62%
Primary: numpy
Secondary: matplotlib, sklearn, pandas, keras.utils (for preprocessing)
Prathamesh Mandke - mandkepk97@gmail.com
This project is licensed under the MIT License - see the LICENSE.md file for details.