This repository has 4 different tasks performed with LeNet5 on the Cifar10 Dataset

• Cifar 10 Image Classification
• Cifar 9 Image Classification (dropped the last class)
• Cifar 9 where Truck and Automobile are treated as the same class
• Cifar 9 where Truck and Bird are treated as the same class

LeNet5 was introduced by Yan LeCun back in 1998. It's a form of a convolutional network originally intented for the classification of the MNIST Dataset (Handwritten Digits).

Here I made a few changes to the network:

• Activation Functions: ReLu
• Kernel Size: 3x3

Training went on for 10 epochs with a batch size of 32

• Loss Function: Categorical Cross Entropy
• Optimizer: SGD

Both loss and accuracy changed only very slighly. The training set is relatively small (In the world of Deep Learning that is). Theortically the network should perform better if it has less classes to distinguish, but in this case it doesn't make a big difference.

The loss went down slightly which indicates a better performance of the model, but hte difference is so small it's negelectible.

If we look at the accuracy we can see, that it went up by about 3%. Trucks and Automobiles should have a lot of visual features in common so up to a certain point it makes sense to treat them as the same class and as we can see, it didn't worsen the network, but it didn't improve it that much either.

It is notworthy though that this one had the highest accuracy of all 4 experiments.

Now Trucks and Birds should really have nothing in common, but surprisingly the Loss went down slighty and even the accuracy is higher when compared to the Cifar9 version but still below the one, where Trucks were labeled as Automobiles.

I would've expected this one to perform the worst since it doesn't make sense at all to label Trucks as Bids, but surprisingly it performed on par if not even slightly better than the rest.

The implementation was done using Keras with TensorFlow as the backed. Everything you need will be installed via the requirements.txt file

pip3 install -r requirements.txt

You should preferably use a virtual environment.

The Cifar10 dataset will be downloaded when you first run one of the python scripts

• Keras
• TensorFlow
• Matplotlib
• Cifar10
• LeNet5
• Yan LeCun
• MNIST