In the previous article, someone raised the question in a comment about whether generative adversarial networks could be used to simulate shooting covers. This was previously tested in a simple experiment, but the results were not ideal.
The unsatisfactory results stem from the implementation of general generative adversarial networks, which essentially involves a transformation or mapping from a latent space to a sample space. When learning high-dimensional data, it often exhibits certain flaws; intuitively, it usually fails to effectively simulate concepts of cycles and periods.
For example, if we attempt to learn a clear distribution like this:
However, the actual result after training looks like this:
Moreover, this cannot be further optimized.
Originally, we hoped it would turn out beautifully:
But in reality, it:
Furthermore, if we assume that all things can be divided into eight categories, then in an infinitely large space, a limited vocabulary can be clustered into eight groups. Using eight Gaussian clusters for representation would look something like this:
The left side can be understood as various things mixed together, while the right side can be seen as a similar classification into eight trigrams, each conforming to a Gaussian distribution and represented by different colors indicating categories.
The ideal classification process should resemble this:
However, the actual process during GAN generation looks like this:
Therefore, the results it generates can only be like this, and the classification effect is clearly not ideal.
This also gives us some meaningful insights; although adversarial performance can be good in certain tasks, it inherently has many issues. Abstracted to reality, just like war cannot solve all problems.
If we extend this further, we find something even more interesting: generative adversarial networks are like a divination, which not only has adversarial implications but also embodies exchange, migration, and transformation. It seems that GANs are better suited for tasks like style transfer and variation, rather than creation.
Through certain methods, the aforementioned issues can be overcome, but that requires more complex research. Here, I merely mention it briefly; interested researchers can reach out for discussion.