Impact of Task Similarity and Training Regimes on Cognitive Transfer and Interference

Nicholas Menghi, Simone Viganò, W. Jeffrey Johnston, Salma Elnagar, Stefano Fusi, Christian F. DoellerbioRxiv 2025 · DOI

Abstract

Learning depends not only on the content of what we learn, but also on how we learn and on how experiences are structured over time. To investigate how task similarity and training regime interact during learning, we trained participants on spatial and conceptual learning tasks that shared either similar or distinct underlying structures, using either interleaved or blocked regimes. Interleaving the two tasks hindered performance when their structures were similar, compared to when they were different. In contrast, blocked training produced the opposite effect: it improved performance and facilitated transfer across similar tasks. This effect, however, emerged only when participants first learned the conceptual task, followed by the spatial task, suggesting an asymmetric interaction between task order and structural similarity. We also replicated our results using a neural network model, providing converging evidence for the computational principles governing the interplay between training regime and structural similarity in multi-task learning.

Introduction

Humans possess a remarkable capacity for "continual learning," the ability to acquire new skills or information without overwriting previous knowledge. In contrast, artificial neural networks often struggle with "catastrophic forgetting," where new learning erases prior representations. A central question in cognitive science is how the underlying structure of tasks—their similarity or difference—interacts with the training regime (the temporal presentation of information) to influence learning outcomes.

While conventional pedagogical wisdom often favors "interleaved" training (mixing different tasks) for better long-term retention, this research explores a counter-intuitive hypothesis: when tasks are highly similar, interleaving may induce confusion and interference. Conversely, "blocked" training (mastering one task before starting the next) might allow the first task to serve as a structural scaffold for the second.

Methodology

The researchers employed a dual-pronged approach, combining human behavioral experiments with computational modeling.

Human Behavioral Study

180 participants were tasked with classification problems across two domains:

  • Conceptual Task: Predicting if a seed "grows" or "dies" based on the opacity of "sun" and "water drop" icons.
  • Spatial Task: Predicting the same outcome based on (x,y) coordinates on a map.

The study manipulated two key variables: task structure (Same vs. Different) and training regime (Interleaved vs. Blocked).

Experimental design illustrating conceptual and spatial tasks, structural similarity conditions, and training regimes.

Computational Modeling

A Recurrent Neural Network (RNN) was used to simulate the human results. By manipulating the "cosine similarity" of input features, the researchers investigated how representational overlap affects learning trajectories in blocked versus interleaved setups.

Key Insights and Results

The study revealed a complex interplay between how tasks are presented and how similar they are.

The Similarity Trap and Blocked Scaffolding

A primary finding was the "similarity trap" in interleaved learning. When tasks shared the same structure, interleaving them led to significantly worse performance compared to when structures were different ($p = 0.005$). This suggests that the brain attempts to use shared compositional representations, leading to interference when tasks are mixed.

In contrast, blocked training acted as a bridge. Participants in the blocked group performed significantly better overall during both training ($t(168) = 6.693, p < 0.001$) and testing ($t(168) = 3.488, p < 0.001$). For these participants, a shared structure was an advantage; the first task provided a template that accelerated learning of the second.

Performance accuracy across training, test, and generalization phases, categorized by training regime and task structure similarity.

Asymmetric Transfer

Interestingly, the benefit of blocked training was asymmetric. Transfer was only observed when participants moved from the Conceptual to the Spatial task ($p = 0.034$). The authors hypothesize that conceptual rules are more abstract, requiring information compression. Once formed, this abstract rule is easily mapped onto concrete spatial layouts, whereas the reverse mapping is more cognitively demanding.

Neural Network Alignment

The RNN model successfully replicated these human behavioral patterns. It demonstrated that blocked training is superior for similar tasks because it allows the network to stabilize in a useful representational space without the constant disruptive "push" from a competing interleaved task.

Discussion and Future Directions

While blocked training improved performance on previously seen ("old") configurations, it did not significantly enhance generalization to entirely "new" configurations compared to interleaved training. This suggests the benefit is primarily driven by faster mapping and improved memorization.

Future research should explore the "abstract-to-concrete" asymmetry hypothesis and investigate how these training regimes affect long-term memory consolidation over multiple days. Furthermore, applying these "curriculum learning" principles to large-scale AI models remains a critical challenge for mitigating catastrophic forgetting in machine learning.

Summary updated

Abstract

Abstract Learning depends not only on the content of what we learn, but also on how we learn and on how experiences are structured over time. To investigate how task similarity and training regime interact during learning, we trained participants on spatial and conceptual learning tasks that shared either similar or distinct underlying structures, using either interleaved or blocked regimes. Interleaving the two tasks hindered performance when their structures were similar, compared to when they were different. In contrast, blocked training produced the opposite effect: it improved performance and facilitated transfer across similar tasks. This effect, however, emerged only when participants first learned the conceptual task, followed by the spatial task, suggesting an asymmetric interaction between task order and structural similarity. We also replicated our results using a neural network model, providing converging evidence for the computational principles governing the interplay between training regime and structural similarity in multi-task learning.
BibTeX
@misc{menghi2025impact,
  author = {Nicholas Menghi and Simone Viganò and W. Jeffrey Johnston and Salma Elnagar and Stefano Fusi and Christian F. Doeller},
  title = {{Impact of Task Similarity and Training Regimes on Cognitive Transfer and Interference}},
  year = {2025},
  doi = {10.1101/2025.09.22.677779},
  howpublished = {bioRxiv},
}
Figure