CNN and Transfer Learning

- 1 min

Alt text

At the heart of the hype of Neural Networks is Convolutional Neural Network, a cutting edge technology that enables computers to identify objects in images. This paper attempts to build a baseline CNN (test accuracy of 43%) to visually identify 20 different food items using the Food101 data. Moreover, the paper builds a custom model (test accuracy of 55%) as well as pre-trained Resnet and VGG16 models to better predict the food images. Through this exploration, the team learned that pre-trained models perform best with weight-freezing (test accuracy of 68% for ResNet and 70% for VGG-16) and without, weight-freezing (test accuracy of 63% for ResNet and 5% for VGG-16), models do not promise sta- ble accuracy results. Moreover, the experiment made it clear that fully connected layers do play a significant role in creating CNN.