Same Movements, Different Skills: How Context Shapes Learning in Humans and Neural Networks
This paper demonstrates that changing performance contexts can cause identical movement sequences to be learned as distinct skills rather than integrated representations, a phenomenon observed in both human learners and recurrent neural networks when strong contextual inputs coincide with high initial weight scales.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
We often assume that learning is a matter of stacking experiences into a single, growing pile of knowledge. If you learn to tie your shoes, you can do it in the kitchen or on the porch; if you learn a melody on the piano, you can play it with your eyes open or closed. This intuition suggests that the brain treats the core action—the sequence of movements—as a unified skill, separate from the surroundings in which it happens. In the world of motor learning, scientists have long debated whether this is true. Does the brain bundle different experiences together to create a general ability, or does it keep them separate, treating the same movement as a different task depending on where and how it is performed? The answer matters because it reveals how we adapt to new situations and why we sometimes struggle to transfer a skill from one setting to another.
A team of researchers at Princeton University set out to test the limits of this generalization. They asked a simple but profound question: if a person practices the exact same sequence of finger movements, but does so while the surrounding environment keeps changing, will they learn one skill or two? To find out, they turned to both human volunteers and computer models. They designed a task where the physical action never changed, but the "story" told around it did. In one version, the sequence looked like musical notes scrolling across a screen, accompanied by tones. In another, the same sequence moved a character through a digital maze. For the computer models, they built simple networks that mimic how neurons pass signals to one another, allowing the researchers to peek inside the "brain" of the machine to see how it organized the information.
The results challenged the idea that identical movements always lead to identical learning. When the computer networks were set up with a specific kind of flexibility and exposed to strong environmental cues, they stopped treating the two versions of the task as the same thing. Instead of building one shared memory for the finger sequence, they built two separate ones. When the context switched from music to maze, the network stumbled, as if it had to relearn the sequence from scratch. This "switching cost" was a clear sign that the machine had split the skill into two distinct parts. Crucially, this split only happened when two conditions were met: the network needed to be flexible enough to reshape its internal connections, and the environmental cues needed to be strong enough to grab its attention. If either condition was missing, the network lumped the experiences together and generalized smoothly.
The human participants showed a strikingly similar pattern. People practiced the same eight-finger sequence on a keyboard, sometimes staying in one context and sometimes alternating between the musical and maze versions. Those who alternated contexts did not show a single, smooth line of improvement. Instead, their performance curves broke apart. When they switched from the maze to the music, their speed dropped significantly, as if they were starting over. When they switched back, they dropped again. This happened even though the finger movements were identical and the context provided no useful information for solving the task. The brain, it seemed, had decided that the context was part of the skill itself. By changing the setting, the researchers had inadvertently turned one skill into two.
To understand what was happening under the hood, the researchers used a mathematical framework to model the learning process. They tested whether the data fit a model where learning was shared across contexts, a model where the contexts influenced each other slightly, or a model where they were completely separate. The best fit for the human data was the split model, suggesting that people were indeed learning two distinct versions of the same sequence. The computer models confirmed this, showing that the ones that split their learning also showed the largest gaps in their internal representations. The study suggests that the brain does not just record movements; it records the entire situation. When the situation changes, the brain may decide to store the movement as a new, separate entity rather than updating the old one.
This finding reshapes how we think about practice and mastery. It implies that the boundaries of a learned skill are not defined solely by the physical movements required to perform it. Instead, the context in which we learn plays a powerful role in how that skill is stored and retrieved. If you learn a sequence of notes while playing a specific song, you might not be able to use those same finger movements as easily when playing a different song, even if the notes are identical. The brain treats the surrounding context as a key part of the instruction. The researchers found that this separation is not a mistake or a flaw, but a feature of how learning systems organize experience. Whether we lump experiences together or split them apart depends on how much weight we give to the environment and how flexible our internal systems are in adapting to it.
The study also highlights the delicate balance required for generalization. In the computer simulations, the networks only split their learning when they were given the freedom to reorganize their internal structure and when the environmental cues were loud enough to demand attention. Without that flexibility, they clung to a single, shared memory. Without strong cues, they ignored the context and generalized freely. This suggests that the human brain operates in a similar way, constantly weighing the importance of the task against the importance of the setting. When the setting feels significant, we learn context-specific skills. When it feels irrelevant, we learn general ones.
Ultimately, the work reveals that learning is not a simple accumulation of facts or movements. It is a dynamic process of deciding what belongs together and what must be kept apart. The same sequence of finger taps can be one skill or two, depending on the story the brain tells itself about where those taps are happening. This insight helps explain why we sometimes struggle to apply what we have learned in the classroom to the real world, or why a skill learned in one environment feels foreign in another. The brain is not just learning the movement; it is learning the world in which the movement takes place. And in doing so, it creates a rich, complex map of experience where the same action can have many different meanings.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.