Common regularization techniques in ML - Data augmentation

 

 

 

Data augmentation is a technique used in machine learning and computer vision to artificially increase the size and diversity of the training dataset by applying a set of transformations to the existing data.

The main goal of data augmentation is to introduce variations in the training data so that the model can learn and generalize better. By generating additional examples, data augmentation helps to overcome the limitation of having a limited amount of labeled data available for training. This technique is particularly useful when the size of the training dataset is small, as it allows the model to learn from a more diverse set of samples.

Here are some common transformations used in data augmentation:

1. Cropping: Randomly selecting a part of the image, discarding the rest. This helps the model learn to focus on important features within the image. For example, when training a model to recognize faces, cropping can be used to extract different facial regions (eyes, nose, mouth) from the original images.

2. Rotation: Rotating the image by a certain angle. This can help the model learn rotational invariance, making it more robust to images that are not perfectly aligned. For instance, when training a model to classify objects, rotating the images can simulate different viewing angles for each object

 

3. Scaling: Resizing the image to a different size, either larger or smaller. Scaling can help the model learn to recognize objects at different scales and improve its ability to generalize. For example, when training a model to detect objects in varying distances using a single camera, scaling can simulate the different sizes of objects in the real world.

4. Flipping: Mirroring the image horizontally or vertically. This transformation helps the model learn orientation invariance. It helps the model recognize that an object can appear flipped or mirrored in different images. For instance, when training a model for character recognition, flipping can generate new training samples, such as turning a "D" into a "b" or a "p".

5. Noise Injection: Adding random noise to the image. This can make the model more robust to different types of noise present in real world data. For example, when training a model for image denoising, adding random noise during augmentation can simulate the noise that might be present in the actual images.

These are just a few examples of transformations used in data augmentation. There are many other techniques and combinations of transformations that can be applied, depending on the specific task and dataset. The goal is to generate new and diverse training samples that capture different variations and scenarios that the model may encounter during testing or deployment.