Self-reinforcing learning can help us to understand new things, but also to reinforce false beliefs. Neuroscientist Franziska Bröker is investigating whether people can continue to learn something on their own, i.e. develop learning independently without feedback.

Tübingen, October 18. 2024 Imagine a child seeing sheep and goats on a farm for the first time. A parent explains which animal is which, and after a few pointers, the child learns to tell the two apart. But what happens if the child visits the farm again a few weeks later and doesn’t get this support? Will it still be able to remember the features that distinguish goats from sheep? Neuroscientist Franziska Bröker has investigated precisely this: How both humans and machines learn without guidance – comparable to a child left to its own devices and discovering the world around it. Her findings: learning without help can sometimes help us, but under certain circumstances it can also make us look pretty stupid.

In machine learning, algorithms are able to sift through large amounts of data and recognise precise patterns without receiving direct feedback from outside. Humans, on the other hand, often find it difficult to learn without feedback, as they can tend to internally reinforce faulty assumptions if no one corrects their mistakes.

Neuroscientific research has shown that unsupervised learning, i.e. learning without external feedback, is particularly successful if the self-made assumptions already correspond relatively well to a solution. For more complex tasks, however, such as learning languages or a musical instrument, feedback is essential in order to avoid mistakes. It is therefore not so much a question of whether learning without feedback works in principle, but rather in which situations it makes sense. It is therefore crucial to understand the framework conditions under which independent learning is effective in order to develop better teaching methods and algorithms.

Feedback as the key to expert knowledge

Laboratory studies show that unsupervised learning sometimes works in theory and sometimes not. But what does it look like in practice, especially when it comes to acquiring specialist expertise? Let’s take radiologists as an example: They receive a lot of feedback in the first few years of their training, but at some point they are left to their own devices. Research suggests that experience alone is often not enough to develop reliable expertise. Rather, it often only indicates a certain seniority than actual ability. Cognitive biases, such as the confirmation bias, can hinder further learning success as we tend to rely on information that reinforces our existing beliefs rather than reflecting on them. Regular feedback is therefore crucial to ensure that learners stay on track and continue to develop their skills.

As described in the example with the radiologists, unsupervised learning is often driven by self-reinforcement mechanisms in which learners rely on their own predictions rather than seeking outside support. This problem is particularly evident in the development of specialised knowledge. Without external correction loops, incorrect assumptions persist and are even reinforced – regardless of whether they are correct or not. These learning traps occur preferably when we stop being motivated to look for alternative solutions or information and instead rely on our own, often erroneous and entrenched assumptions.

It’s all about striking the right balance

Research on the topic shows that self-reinforcement can be both an advantage and a disadvantage in the context of independent learning without assistance. The principle of self-reinforcing learning can help us to understand new things, but it can also lead to us becoming entangled in false beliefs. Unsupervised learning has a lot of potential, but it is not a miracle cure – it depends heavily on how well our knowledge, our inner motivational impulses, our own learning ability and the task structure harmonise.

Future studies should therefore investigate how self-reinforcement through personal exploration and discovery and, on the other hand, learning through external feedback can work well together, especially in real learning situations. This is not just about machine learning, but above all about the learning process in humans. In this way, new teaching methods can be developed that promote lifelong learning and help to avoid incorrectly learnt assumptions.

Original publication:

Bröker, Franziska et al. (2024): Demystifying unsupervised learning: how it helps and hurts.
Trends in Cognitive Sciences, doi:10.1016/j.tics.2024.09.005

Further Informationen

(https://www.kyb.tuebingen.mpg.de/794392/news_publication_23575937_transferred)

ImageSource Pexels Christina u Peter