hedzd / Fruit-360-classification

Classifying fruits on the Fruit-360 dataset by creating a fully connected artificial neural network from scratch.

Fruit-360-classification

In this project, a fully connected artificial neural network is implemented from scratch.

Neural network architecture and details

This ANN was implemented to classify 4 classes of fruits. Feedforward algorithm was implemented in vectorized form using softmax as activation function for each layer. Back propagation was implemented in both iterative and vectorized forms with sum of squared errors (SSE) as cost function. Stochastic Gradient Descent algorithm was used to train the network.

Additional parts included:

• Hyperparameter tuning
• Improving SDG using momentum algorithm
• Adding more classes of fruits and hyperparameter tuning
• Using softmax as output layer's activation function

Dataset

Kaggle 360-Fruits dataset was used.
A feature extraction and size reduction technique was used on train and test dataset to simplify the problem.