Neural networks with extensively deep architectures typically contain millions of parameters, making them both computationally expensive and time-consuming to train.
In this tutorial, we'll achieve state-of-the-art image classification performance using DenseNet, initially with a single hidden layer. We systematically tune the number of neurons in the hidden layer and train our model on a benchmark image classification dataset. This study shows that building deeper neural networks is not always necessary; instead, it is more important to focus on the correct number of neurons in each layer.
For a more detailed breakdown of the code, check out the full tutorial on the blog.