Supervised Approaches to CIFAR10

Supervised machine learning is a widely used form of artificial intelligence. There are plenty of ways to approach supervised learning: some of them being Neural Networks, Convolutional Neural Networks and Residual Networks. In this repository we develop an in depth analysis of the difference between these on the CIFAR10 classification task.

CIFAR10 dataset. Sample of images with corresponding labels.

Deep Linear NN

Testing Loss: 1.516 Testing Accuracy: 0.462

Deep CNN

No Data Augmentation

Testing Loss: 1.109 Testing Accuracy: 0.675

Data Augmentation

Testing Loss: 0.876 Testing Accuracy: 0.708

Pre-trained ResNet18

No resizing

Testing Loss: 0.740 Testing Accuracy: 0.820

Resized (224,224)px images

Testing Loss: 0.182 Testing Accuracy: 0.952

Credits

• m4mbo - Code
• LMH summer program on 'AI and ML: Advanced Applications' - Theory