Data Science

Training a model on more data can lead to more accurate $\hat{y}$
This is especially true for convolutional networks
We won't always have the opportunity to go and get more images
Therefore, we can take our images, transform them, and add those new images to our set of images
This process is referred to as data augmentation

Data augmentation can refer to many different types of image transformation techniques
The following techniques are some popular ones:
- Mirroring
  - This refers to flipping an image to create a new image
  - For most computer vision tasks, mirroring an image will still preserve the same class of image
  - For example, flipping a picture of a cat will still give us a picture of a cat
- Random cropping
  - This refers to cropping certain areas of the image to create a new image
  - For most computer vision tasks, this will still preserve the same class of image as well
  - However, we need to be more careful with applying this method, compared to mirroring
Other techniques include rotation, shearing, and local warping
However, these augmentation techniques are less common, simply due to an additional level of complexity

A third type of data augmentation that is commonly used is referred to as color shifting
This refers to creating new images by adjusting the RGB channels
Specifically, we distort the R, G, and B color channels by adding fixed values to each channel
In practice, these fixed values are drawn from some probability distribution
For example, we can decrease the red and green channels and increase the blue channel to make the image yellower
We may want to do this to create an image with the effect of having more sunlight
This image would have a new effect, but maintain the same identity
By doing this, our learning algorithm becomes more robust to changes in colors of our images
Typically, we implement distortions during training

colorshifting

We can implement color shifting using principal component analysis
Specifically, PCA color augmentation is designed to shift values that are more present in the image
For example, images with heavy red values and minimal green values will have their red values altered the most
The specific mechanism relies on the PCA algorithm to find the relative color balance of a given image
The AlexNet paper goes into this subject in more detail

Data augmentation can refer to many different types of image transformation techniques
The following techniques are some popular ones:
- Mirroring
- Random cropping
- Color shifting
Other techniques include rotation, shearing, and local warping
However, these augmentation techniques are less common

Transfer Learning

Object Localization