What’s next to improve the model accuracy of Convolutional Neural Networks (CNNs)

Since the success of AlexNet in 2012, deep convolution neural networks have become the dominating approach for some areas of computer vision. Many recent studies are focusing on proposing new architecture to improve model accuracy. Most of the CNN architecture introduces new techniques from the model architecture level. He et al. [1] suggested that only changing model architecture is not responsible for improving model performance. Training procedures and Network tweaks also add value to gaining more model accuracy.

Image is taken from https://wallpapercave.com/

In this blog post, I will try to include their findings so that in the future, I can look at this list and try one by one to improve my model performance.
I can start with an example project I recently implemented for my company. The task is straightforward in the computer vision domain, and it is image classification. I was asked to implement image classification to recognise this image is a child product or not. It is a binary class classification. As I previously knew, many CNN architectures are already available to implement this task.

I followed this step to set up my baseline model —

• I looked into the dataset and I noticed there were 2263 sample images for both categories. I am not sure for image classification…