Background

The goal in the internship was to code and present an AI; but, AI is a broad field with many subcategories. The one that I dealt with was Convolutional Neural Networks or CNNs. These take in an image and process parts of the image based on color and texture, or “weights.”

These weights are usually used for visual purposes such as object recognition inside the image. Then, they divide up each section of the image further and take the weights of each part, in steps called convolutions.

​

​

​

​

​

​

​

​

​

​

​

Each AI has different numbers and types of layers, each performing a unique function. Usually, the more layers the AI has, the better the performance will be but the lower the efficiency will be. ​

Good quality CNN models, with varying numbers of layers, are trained (most commonly on the sophisticated and widely used ImageNet dataset) and uploaded to the internet for others to use. These models are referred to as pre-trained models, which can be fine-tuned, or re-trained on other datasets,

for specific purposes.

​

There are various methods of fine-tuning these pre-trained models and various languages that can be used to train them. As I was new to the process of fine-tuning, I chose the most user-friendly of these: MatLab's deep network designer. It allows the user to import the model and setting directly adjust training settings.