Abstract
Transfer learning, the reuse of newly acquired knowledge under novel circumstances, is a critical hallmark of human intelligence that has frequently been pitted against the capacities of artificial learning agents. Yet, the computations relevant to transfer learning have been little investigated in humans. The benefit of efficient inductive biases (meta-level constraints that shape learning, often referred to as priors in the Bayesian learning approach), has been both theoretically and experimentally established. Efficiency of inductive biases depends on their capacity to generalize earlier experiences. We argue that successful transfer learning upon task acquisition is ensured by updating inductive biases and transfer of knowledge hinges upon capturing the structure of the task in the inductive bias that can be reused in novel tasks. To explore this, we trained participants on a non-trivial visual stimulus sequence task (Alternating Serial Response Times, ASRT); during the Training phase, participants were exposed to one specific sequence for multiple days, then on the Transfer phase, the sequence changed, while the underlying structure of the task remained the same. Our results show that beyond the acquisition of the stimulus sequence, our participants were also able to update their inductive biases. Acquisition of the new sequence was considerably sped up by earlier exposure, but this enhancement was specific to individuals showing signatures of abandoning initial inductive biases. Enhancement of learning was reflected in the development of a new internal model. Additionally, our findings highlight the ability of participants to construct an inventory of internal models and alternate between them based on environmental demands. Further, investigation of the behavior during transfer revealed that it is the subjective internal model of individuals that can predict the transfer across tasks. Our results demonstrate that even imperfect learning in a challenging environment helps learning in a new context by reusing the subjective and partial knowledge about environmental regularities.
Introduction
Transfer learning—the ability to take what you learned in one context and apply it to another—is a cornerstone of human intelligence. While artificial agents often struggle with this, humans do it effortlessly. However, the exact computations that allow us to do this are poorly understood. This paper by Székely et al. explores how humans reshape their high-level mental frameworks—their inductive biases—to become more efficient learners in new, but related, environments.
The authors address a specific gap in cognitive research: the assumption that humans learn a "ground truth" version of a task. In reality, human learning is often imperfect and highly individual. This study asks: How do these idiosyncratic, subjective internal models shape the inductive biases we transfer to new tasks? Do we transfer the "correct" rules of the world, or do we transfer our own personal (and potentially flawed) version of those rules?
Methodology: ASRT and Cognitive Tomography
The researchers used a non-trivial sequence learning task called Alternating Serial Response Times (ASRT). In this task, participants see a sequence of visual stimuli where some elements follow a deterministic pattern (the "structure"), while others are random. The experiment spanned 10 sessions, including a long Training phase (Task 1), a Transfer phase (Task 2, which used a permuted version of the same structure), and an Alternation phase where both tasks were mixed.
To "see" inside the participants' heads, the authors employed a computational framework called Cognitive Tomography (CT). Using an Infinite Hidden Markov Model (iHMM), the CT framework infers a participant's internal model solely from their response times (RTs). The logic is simple: if a participant expects a stimulus based on their internal model, they react faster. These individual CT models were compared against a simple Markov model (a baseline that only looks at the immediate previous step) and the Ground Truth (the actual logic of the experiment).
Key Results and Insights
Reshaping the Bias
Humans start with a simple "Markovian" bias, expecting events to only depend on the immediate past. Over eight days of training, participants successfully "abandoned" this simple bias and built complex internal models that matched the task's hidden structure. This was quantitatively shown by the increasing cross-entropy between the CT models and the initial Markov bias from Day 1 to Day 8.
Subjective Transfer is King
When moving to Task 2, participants didn't just transfer the "correct" ground truth. Instead, their performance on the new task was best predicted by a transformation of their own subjective, imperfect model from Task 1. The permuted subjective internal model explained significantly more variance in Task 2 response times than the permuted Ground Truth model (mean $r^2$ of 0.075 vs 0.019). We transfer our interpretation of the world, not the world itself.
The Mental Inventory
Participants didn't overwrite Task 1 when they learned Task 2. On Day 10, they were able to switch back and forth between the two models based on the stimuli, showing that humans build an "inventory" of internal models and recruit them as needed. Furthermore, individuals with higher "Learning Strength" on the first task—defined as the predictive power of the CT model beyond a simple Markov prior—were significantly better at transferring that knowledge to the second.
Discussion and Future Work
While Cognitive Tomography provides a powerful "snapshot" of internal representations, the study notes limitations in tracking rapid, trial-by-trial changes due to the data requirements of the iHMM. Future work could incorporate other data, like eye-tracking, to provide a richer picture of predictive probabilities. Additionally, investigating if training on a more diverse set of tasks leads to even more robust and abstract inductive biases remains an open question. Ultimately, these results demonstrate that even imperfect learning in a challenging environment helps learning in a new context by reusing subjective and partial knowledge about environmental regularities.
Summary updated