Artificial data synthesis vs Data augmentation

Course Q&A Deep Learning Specialization Structuring Machine Learning Projects
1

I have two questions.

  1. Does the artificial data synthesis generate fake data based on the data provided in the training phase? As a result, it may contribute to the overfitting issue.

  2. What is the difference between data augmentation and artificial data synthesis, and which is better to use? I have some confusion about them.

2

Hi there,

Artificial data synthesis maybe be creation of data that is similar to the the available data, modelling, generation, artwork etc. The data will be the same as training set but not exactly and therefore will add learning value to the model. The subject of overfitting is not straightfoward and you can never completely separate data leakage but for the purpose of conventional model training I dont think its an issue with data synthesis.

Data augmentation is when the existing data is tranformed by rotation, flipping, cutting, cropping etc. it has similarities with the training data but its not the exactly the same so it will also add learning value.

Which is better to use? The artificial data depending on the kind of synthesis could include very hard to find examples, to train the model, cases which you may not easily find. The comparison indicator is not straighforward as many factors are involved though.

1 Like
3

Hi @Areeg_Fahad

  1. Does the artificial data synthesis generate fake data based on the data provided in the training phase?

Yes, artificial data synthesis typically generates new data based on the patterns and characteristics of the data provided in the training phase. This can include methods such as adding noise, rotating or flipping images, or creating new samples by combining existing samples. The goal is to increase the size of the training set and introduce more variability, which can help to improve the generalization of the model. However, if the synthetic data is not diverse enough, or if the method used to generate the synthetic data is not appropriate for the task, it can contribute to overfitting.

  1. What is the difference between data augmentation and artificial data synthesis, and which is better to use?

Data augmentation and artificial data synthesis are similar in that they both aim to increase the size of the training set and introduce more variability in the data. However, there is a subtle difference between the two.

Data augmentation typically involves applying simple, deterministic transformations to the existing data, such as flipping, cropping, or rotating images. The goal is to increase the size of the training set by creating new, slightly different versions of the existing data.

Artificial data synthesis, on the other hand, typically involves creating new data from scratch, often by combining or modifying existing data. This can include methods such as adding noise, creating new samples by combining existing samples, or using generative models such as GANs.

Which is better to use depends on the specific task and dataset. Data augmentation can be a simple and effective way to improve the generalization of the model, while artificial data synthesis can be more powerful but also more complex. It’s often a good idea to try both and see which works best for your specific problem.

It’s also worth noting that both data augmentation and artificial data synthesis can be used in combination with other regularization techniques like dropout or L1/L2 regularization to prevent overfitting.

Regards
Muhammad John Abbas

1 Like