No Panic for AI

When everyone is talking about AI and worrying about AI, especially worrying about losing jobs due to AI, I strongly believe that the spring of AI has yet to come. Please, no panic for it.

Am I joking? Haven’t I seen the emergence of products like chatting robots or the revolution of deep learning?

The chatting robots that are based on generative AI are basically trained by text. There are some developments on audio and images, but not that much for video, as the size of video is still too big for most machine learning platforms.

In addition, these robots only generate answers that look good. I mean, the flow of language is smooth and without grammar mistakes, but the content is still of low quality. This is because it is still based on a strong assumption that all information in the training dataset is correct. To filter out incorrect information in the training dataset requires tedious human labor, which means the development of AI products itself requires large human resources who don’t have to know the underlying theories.

Therefore, please don’t panic about losing jobs, even when you don’t know much of the technique.

Have you got experience comparing texts and videos on your computer? The size of a two-hour movie will be much larger than a huge number of text-based novels. This is why I think the techniques are still vulnerable when it comes to analyzing videos.

The annotation of images or videos will be even more challenging.

In a lot of situations, if we want to detect unusual events, for example, dangerous driving in traffic surveillance, it is still an unsolved problem to collect a dataset with a sufficient number of those events. Simulation is a direction, but there is no real research in this direction so far.

Process mining looks fantastic. But do you know that the log data in information systems are not in the format that can be used directly by the algorithms? There is a tedious process to prepare data initially and there are many ambiguous matters.

Many papers that use deep learning techniques on small datasets claim that the proposed techniques are more advanced than using shallow learning like SVM. But in fact, they simply get good experimental results due to overfitting on the small datasets.

The ambitious proposal of channel sensing in communication channels in 6G is unrealistic at this stage. There is a lack of foundation from machine learning theories to experimental designs in this area. How can it be real around 2030?