Academic project which refers to implementation of Mini-Batch Stochastic Gradient Ascent (SGA). The implementation consists of a MultiLayerPerceptron (MLP) with one hidden layer with M hidden units. The Neural Network was trained upon MNIST and CIFAR-10 datasets and it tried to predict the image.
- Python
- NumPy
- Matplotlib
- cPickle
The MLP was trained with different hyperparameters.
- Learning_rates = = [ 0.01, 0.001]
- λ = [ 0.1, 0.5 ]
- epochs = [ 10, 20, 30]
- HiddenLayers (M) = [100, 200, 300]
- activation_h = [h1. h2. h3]
- batch_size = 200
In MNIST DATA we got accuracy > 85.6% with best the acc = 98.14%.
In CIFAR-10 DATA we got accuracy > 34.5% with best the acc = 45.81%.
You can observe every single run in corresponding Excel file.