This approach (see notebook) investigates whether multiple CNN models can achieve higher classification accuracy than any individual model.
- Classification based on the average class probabilities of models
- Using the mode class for prediction
- A simple CNN can achieve classification accuracy of over 93%
- Combining 3 models improves accuracy around 94.4%
- It takes around 16 seconds per epoch using Colaboratory GPU accelerator and Test accuracy does not improve significantly after the first 20 epochs
- Combining a few more models trained over 20 epochs may further improve classification accuracy in a resonable amount of time.
- Classification accuracy is significantly lower for 4 classes: 'T-shirt/top', 'Pullover', 'Coat', and 'Shirt'.
- Devise alternate methods for combining models
- Increase the diversity of constituent models
- Introduce regularization methods that prevent over-fitting beyond 20 epochs
- Develop a two-phased approach: Predict using a combination of models in the first phase and use a separate model to re-classify examples predicted as 'T-shirt/top', 'Pullover', 'Coat', or 'Shirt