NN-from-scratch

This is the code for a fully connected neural network. The code is written from scratch using Numpy, without using any ready-made deep learning library. In this, classification is done on the MNIST dataset. It is generalized to include various options for activation functions, loss functions, types of regularization, and output activation types.

#ann#backpropagation#deep-learning#fully-connected-deep-neural-network#fully-connected-neural-network#gradient-descent

View Repository on GitHub ↗

REPOSITORY OVERVIEW

Live repository statistics

★ 1Stars

⑂ 0Forks

◯ 0Open issues

◉ 1Watchers

31/100

OPENREPOHUB HEALTH SIGNAL

Limited signals

A transparent discovery signal based on current public GitHub metadata.

Recent activity35% weight

Community adoption25% weight

Maintenance state20% weight

100

License clarity10% weight

Project information10% weight

This score does not audit code, security, maintainers, documentation quality, or suitability. Verify the repository and its current documentation before adoption.

README preview

Fully Connected Neural Network from scratch

This is the code for a fully connected neural network. It is written from scratch without using any ready-made deep learning library. All the matrix operations are done using the numpy library. In this, classification is done on the MNIST dataset. We can give any number of hidden layers and any number of neurons in hidden layers as inputs. It is generalized to include various options for activation functions, loss functions, types of regularization, and output activation types. The following are the options available (text should be used as it is written below while giving the arguments for calling the functions):

Available Options (Type as given below while giving inputs as arguments):

Hidden layer activation function: sigmoid, tanh, ReLu (default output activation is sigmoid)
Output activation function: sigmoid (default), softmax (Boolean: True/False)
Cost (Loss function): Quadratic, CrossEntropy
Regularization: l1, l2

Above 97% classification accuracy can be achieved by using [784,100,10] architecture, CrossEntropy loss, l2 regularization, 30 epochs.

Although, MNIST dataset is used in the code, it can do classification for any dataset. Number of neurons in the input layer and the output layer has to be set as per the dataset.