Can MLP neural networks identify Simpsons characters?

This small project explores how well a Multilayer Perceptron (MLP) can classify grayscale images of Simpsons characters.

Approach

• Built a baseline MLP with 2 hidden layers, achieving the accuracy of 25.35%

• Tuned hyperparameters (number of layers, number of neurons, learning rate, activation function). The best model with 4 layers, ReLU, and learning rate 0.00647 was retrained on the big dataset (train + validation). This model achieved the accuracy of 29.75% on the test dataset.

• Tested model robustness on augmented images (rotation, zoom, flipping and combined techniques).

Findings

• The model recognizes characters like Homer and Milhouse well, but struggles with Lisa and Krusty

• Backgrounds, clothes, and unusual contexts confuse the model more than facial features

• Accuracy drops on heavily augmented images by ~9%

Used techniques

• Neural networks with Keras

• Hyperparameter tuning

• Data augmentation

• Visualization

Future works

• Improve the model robustness by training on augmented data

• Broaden hyperparameter search

• Explore CNNs or transfer learning for higher performance on image classification