Learning Without Change: Insights from AI Models

The human brain is capable of learning and adapting to new situations using what it has learned in the past. A great example of this is riding a bicycle. It takes a lot of practice to learn this skill, but someone who has experience with skateboarding might pick up bike riding quickly. This is because they can use their balance and coordination skills from skateboarding to help them learn to ride a bike. This ability to apply past knowledge to new situations raises an important question: how much can a learning system use what it already knows? Can it perform complex tasks based only on what it knows without changing its structure?

In humans, the ability to change and adapt in the brain is known as neural plasticity. In artificial intelligence, particularly in models called Artificial Neural Networks, this change corresponds to adjusting weights or parameters in the model. This article looks at how artificial neural networks can learn without needing to change their internal structure.

#Learning in Artificial Neural Networks

Recent studies show that advanced artificial neural networks, particularly large language models (LLMs), can learn new tasks just by looking at a few examples. This ability, known as in-context learning, suggests that these pre-trained systems can perform new tasks using what they already know without needing to adapt their internal structure.

Statistical Learning is a method used to measure how well people can recognize patterns in new data. This method often requires brain plasticity. We’ll look at two statistical learning tasks: Artificial Grammar Learning (AGL) and the Serial Reaction Time Task (SRTT). AGL involves learning rules about sequences, while SRTT measures how quickly a person can react to a series of stimuli.

Even though these tasks differ from the natural language data that LLMs usually process, studies show that LLMs can learn to recognize new statistical patterns without changing their internal structure.

#Previous Research in Neuroscience

Some neuroscience studies align with these findings but usually involve more specific conditions or limited adaptations. Certain models of the brain's hippocampus show that they can generalize patterns learned from examples without further training. These models require special training and specific contexts that closely resemble what they have already learned.

In contrast, our findings show that off-the-shelf LLMs, trained on large amounts of natural language, can generalize complex patterns without any form of adaptation. This suggests new insights into how learning works in the brain and shows how advancements in artificial intelligence can help us learn more about fundamental concepts in cognitive science.

#Testing the Flexibility of Large Language Models

Our experiments focus on the Mistral model, a strong example of a large language model, to see how well it can learn new structures without changing its internal setup. We set up the AGL and SRTT tasks as tasks where the model learns through examples.

We do not make any adjustments to the model's parameters during our testing. Instead, we provide new input step by step, which allows us to see how well the model can learn the structures involved in these tasks.

Similar to classic experiments measuring learning speed, we divided our experiment into three parts: a baseline with random sequences, a structured part with sequences that follow certain rules, and a post-learning part with more random sequences. We want to show learning by comparing how the model performs in the structured part to the baseline.

To make sure our results are reliable, we repeated the experiment with different sets of vocabulary and sequences.

#Results of the Experiments

The results highlight the model's ability to predict outcomes more accurately in the structured part compared to the baseline. We see a clear pattern of improvement, showing that the model can learn from the structured sequences.

Another potential argument could be that the model is merely memorizing sequences. To check this, we ran further experiments by excluding duplicate sequences. Even without repeats, the model still showed learning in the structured part, suggesting that it captures the underlying structure rather than just memorizing specific sequences.

Next, we switched the vocabulary in the middle of the input. This means we used different words that still followed the same structure. The model quickly adapted to this change, indicating that it understands the rules governing the sequences rather than just the words used.

We also tested whether the model learns from random sequences at the beginning of the experiment. Our findings indicate that having a longer baseline of random sequences makes learning more difficult. However, even when we adjust the length of the random phase, the model continues to learn quickly after switching vocabulary.

#The Model's Understanding of Structures

A key feature of the AGL and SRTT is that the model can make different choices at various points. Our analysis shows that the model assigns probabilities to possible next words according to the rules of the structures it is learning. This supports the idea that the model captures the actual patterns of the sequences rather than random associations.

Our approach highlights the model's capability to learn tasks that differ from its original training data without needing further adjustments. Previous studies often required fine-tuning, which involves changing the model to learn new tasks. Our work demonstrates how well pre-trained learning systems can generalize their knowledge to entirely new tasks without any adjustments.

#Connections Between Artificial Intelligence and the Human Brain

There is a growing similarity between how the human brain represents language and how LLMs do the same. This suggests that both systems might respond similarly in learning tasks. The findings from our research suggest that the rich representation generated by LLMs may help them adapt to new situations without needing specific training.

This paper proposes that maybe the brain relies on calculations more than previously thought. It suggests that certain complex tasks can be completed by using existing knowledge rather than needing to change the network's setup. This challenges the traditional view of how information is processed in the brain.

#Conclusion

Our study shows that pre-trained AI systems have a fascinating ability to learn and generalize complex patterns far beyond what they were originally trained on. We believe this offers valuable insights not only into artificial intelligence but also into our own cognitive processes.

The ability of these models to learn without needing adjustments opens up opportunities for further research on how learning and plasticity work in real-world situations. The findings suggest that deep learning models can provide meaningful insights into important ideas from cognitive science and neuroscience, leading to a deeper understanding of both human and artificial learning systems.