The negative transfer problem in machine learning? I can find the definition of negative transfer on Wikipedia:

In behavioral psychology, negative transfer is the interference of the
previous knowledge with new learning, where one set of events could
hurt performance on related tasks.

And a good illustration in this book: The ABCs of How We Learn.

People use
something they have memorized but in the wrong situation. This often
happens because people do not learn to recognize the deep structure of a problem but instead use the obvious surface features as the cue for which solution to apply.

Is it the same as what we talk about in machine learning? What caused the negative transfer? How can we prevent that?
 A: Yes, it's the same concept and fundamental challenge for transfer learning (in terms of both human learning and machine learning. As Lerner mentioned, a good reference will be Characterizing and Avoiding Negative Transfer. Below are quick answers to your question.
To understand the root cause of transfer learning, we need to first understand its objective: given a label-scarce target domain, we aim to borrow relevant samples from a label-rich source domain and build a model that performs better on the target domain than using the target-domain data alone.
Thus, a fundamental task is how to select the relevant samples from the source domain. The selected samples are called either positive or negative samples, depending whether it  improve or degrade the model performance on the target domain, respectively. The inclusion of irrelevant negative samples is called "negative transfer", as it doesn't help and even hurt the learning in the target domain.
Why do we need to choose relevant samples? Because there is discrepancy in the joint distributions between the source and target domains. This is the root cause of negative transfer problem.
I don't think negative transfer is something that can be prevented, but we need to select the relevant samples carefully to minimize its impact. For detailed sample selection methods, you can refer to "Domain Adversarial Network" and "Discriminator Gate" as mentioned in the above literature. Recently, there are also reinforcement learning based methods.
Hope it clarifies you doubts.
