Connecting Context-specific Adaptation in Humans to Meta-learning

Rachit Dubey, Erin Grant, Michael Luo, Karthik Narasimhan, Thomas L. Griffiths

Abstract

Cognitive control, the ability of a system to adapt to the demands of a task, is an integral part of cognition. A widely accepted fact about cognitive control is that it is context-sensitive: Adults and children alike infer information about a task's demands from contextual cues and use these inferences to learn from ambiguous cues. However, the precise way in which people use contextual cues to guide adaptation to a new task remains poorly understood. This work connects the context-sensitive nature of cognitive control to a method for meta-learning with context-conditioned adaptation. We begin by identifying an essential difference between human learning and current approaches to meta-learning: In contrast to humans, existing meta-learning algorithms do not make use of task-specific contextual cues but instead rely exclusively on online feedback in the form of task-specific labels or rewards. To remedy this, we introduce a framework for using contextual information about a task to guide the initialization of task-specific models before adaptation to online feedback. We show how context-conditioned meta-learning can capture human behavior in a cognitive task and how it can be scaled to improve the speed of learning in various settings, including few-shot classification and low-sample reinforcement learning. Our work demonstrates that guiding meta-learning with task information can capture complex, human-like behavior, thereby deepening our understanding of cognitive control.

Introduction

Humans possess an incredible ability for "cognitive control"—the capacity to adapt behavior to changing task demands. A key feature of this flexibility is our use of contextual cues. For example, a bilingual child uses the identity of the speaker (the context) to predict which language (the task) is required before a single word is spoken.

Current meta-learning algorithms, such as Model-Agnostic Meta-Learning (MAML), lack this human-like trait. They typically treat every task as a random draw from a distribution and rely solely on trial-and-error feedback (labels or rewards) to adapt. This makes them less efficient in ambiguous situations where a simple contextual hint could have "primed" the system for success. This research introduces Meta-Learning with Task Information (MLTI) to close this gap by creating a model that uses external context to guide task-specific initialization.

Methodology: Meta-Learning with Task Information (MLTI)

While standard meta-learning learns a single global initialization for all tasks, MLTI introduces a context network ($g_\psi$) to modulate the model based on external cues.

Contextual Priming and FiLM

The context network takes a task-specific cue ($c$) and produces parameters that modulate the base model. To handle high-dimensional models efficiently, the authors use Feature-wise Linear Modulation (FiLM). The context network outputs "shift" ($\beta$) and "scale" ($\gamma$) parameters that linearly transform the activations in the base network's hidden layers.

Two-Step Adaptation

The framework utilizes a nested optimization strategy:

  1. Outer Loop: Learns the global parameters ($\theta$) and the context network weights ($\psi$).
  2. Inner Loop: Uses the context-conditioned initialization and then performs standard gradient descent based on task-specific feedback (the "support set").
The MLTI framework uses task-specific contextual information (c) to learn an initialization for model parameters (φ) before adapting to online feedback.

Experimental Evaluation

The framework was evaluated across three distinct domains to test both its cognitive plausibility and its performance in standard AI benchmarks.

  • Cognitive Modeling: A reproduction of an infant toy-word mapping experiment involving three speakers and two rule sets. The goal was to see if the model's "surprise" mirrored human behavior.
  • Ambiguous Supervised Learning: A modified CelebA dataset where tasks are intentionally ambiguous, requiring context to resolve classification.
  • Reinforcement Learning (RL): Evaluated on MuJoCo and Meta-World benchmarks, where contextual cues like goal positions are vital for fast adaptation.

Key Results

Matching Human Behavior

The authors found that the validation loss of the MLTI model acted as a proxy for infant "looking time" (a measure of surprise). When the context was inconsistent with the task, the model's loss spiked, mirroring human data. MLTI showed significantly lower error in "consistent" vs. "inconsistent" conditions ($3.65$ vs. $6.19$ loss), whereas standard MAML was insensitive to context.

MLTI validation loss qualitatively matches infant "looking time" data, showing sensitivity to inconsistent contexts that standard MAML lacks.

Performance and Efficiency

  • Supervised Performance: On the ambiguous CelebA task, MLTI achieved 77.51% accuracy, outperforming standard MAML (68.61%) and the MAML-CONCAT baseline (75.61%).
  • RL Efficiency: In MuJoCo and Meta-World, MLTI consistently learned faster and reached higher returns than PROMP. The results proved that decoupling context (using it for initialization) from task feedback (using it for adaptation) is more effective than simple concatenation.

Conclusion

MLTI demonstrates that guiding meta-learning with task information can capture complex, human-like behavior and significantly improve learning speed in artificial agents. By treating context as a prior over task structures, models can start much closer to the optimal solution for a specific task. Future work will explore latent context discovery and improving generalization to entirely new types of task descriptors.

Summary updated

Abstract

Cognitive control, the ability of a system to adapt to the demands of a task, is an integral part of cognition. A widely accepted fact about cognitive control is that it is context-sensitive: Adults and children alike infer information about a task's demands from contextual cues and use these inferences to learn from ambiguous cues. However, the precise way in which people use contextual cues to guide adaptation to a new task remains poorly understood. This work connects the context-sensitive nature of cognitive control to a method for meta-learning with context-conditioned adaptation. We begin by identifying an essential difference between human learning and current approaches to meta-learning: In contrast to humans, existing meta-learning algorithms do not make use of task-specific contextual cues but instead rely exclusively on online feedback in the form of task-specific labels or rewards. To remedy this, we introduce a framework for using contextual information about a task to guide the initialization of task-specific models before adaptation to online feedback. We show how context-conditioned meta-learning can capture human behavior in a cognitive task and how it can be scaled to improve the speed of learning in various settings, including few-shot classification and low-sample reinforcement learning. Our work demonstrates that guiding meta-learning with task information can capture complex, human-like behavior, thereby deepening our understanding of cognitive control.
BibTeX
@misc{dubey2020connecting,
  author = {Rachit Dubey and Erin Grant and Michael Luo and Karthik Narasimhan and Thomas L. Griffiths},
  title = {{Connecting Context-specific Adaptation in Humans to Meta-learning}},
  year = {2020},
  eprint = {2011.13782},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2011.13782},
}
Figure