A bigger model
Training your own neural network is an art; you need intuition, some luck, a lot of patience, and all the knowledge and help that you can find. You will also need money and time to either buy a faster GPU, use clusters to test more configurations, or pay to get a better dataset.
But there are no real recipes. That said, we will divide our journey into two phases, as explained in Chapter 5, Deep Learning Workflow:
- Overfitting the training dataset
- Improving generalization
We will start from where we left off in Chapter 4, Deep Learning with Neural Networks, with our basic model reaching 66% validation accuracy on CIFAR-10, and then we will improve it significantly, first to make it faster, and then to make it more precise.
The starting point
The following is the model that we developed in Chapter 4, Deep Learning with Neural Networks, a model that overfits the dataset because it achieves a high training accuracy value at relatively low validation...
