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
In cognitive science, "cognitive control" refers to the human ability to dynamically adapt to task demands based on environmental context. For instance, a child in a multilingual environment can anticipate which language will be spoken based on the person addressing them. This contextual cue acts as a "prior," allowing for near-instantaneous adaptation.
In contrast, modern meta-learning algorithms—designed to "learn to learn"—often operate in a context-blind manner. Popular methods like Model-Agnostic Meta-Learning (MAML) treat each new task as a sample from a distribution, relying entirely on post-hoc feedback (labels or rewards) to adjust parameters. This paper introduces Meta-Learning with Task Information (MLTI), a framework that bridges this gap by using contextual cues to "prime" a model’s initialization, mimicking the context-sensitive nature of human cognitive control.
Methodology: Meta-Learning with Task Information (MLTI)
The core innovation of MLTI is the transition from a single global initialization to a context-dependent one. While standard gradient-based meta-learners learn a fixed initial parameter set $\theta$ for all tasks, MLTI introduces a context network ($g_\psi$). This network processes task-specific information ($c$)—such as a goal vector or a speaker's identity—to generate a customized starting point for the model.
To ensure computational efficiency, the authors employ Feature-wise Linear Modulation (FiLM). Rather than generating all weights for a large neural network, the context network produces "shift" ($\beta$) and "scale" ($\gamma$) parameters. These parameters modulate the activations of the base network, effectively "priming" its state before any task-specific data is processed. The optimization involves an outer loop that updates the context network weights ($\psi$) and global parameters ($\theta$), and an inner loop where the model adapts to specific task feedback starting from the context-conditioned initialization.
Experimental Evaluation
The researchers evaluated MLTI across three distinct domains to test its versatility and cognitive plausibility:
- Cognitive Modeling: Reproducing a developmental psychology study where infants learn object-word mappings from different "speakers." This tested if MLTI could mirror human "surprise" when context and rules are inconsistent.
- Ambiguous Image Classification: Using a modified CelebA dataset where tasks are intentionally ambiguous (e.g., "classify based on two specific attributes"). The model must use a "two-hot" attribute vector as context to disambiguate the task.
- Continuous Control (RL): Benchmarking on MuJoCo (HalfCheetah, Walker2D) and Meta-World (robot manipulation). In Meta-World, the context is the 3D goal position for tasks like "soccer" or "basketball."
Key Results and Discussion
MLTI demonstrated significant improvements in both learning speed and final performance across all benchmarks. In the CelebA classification task, MLTI achieved 77.51% accuracy, outperforming standard MAML (68.61%) and even a baseline that simply concatenated context to the input (75.61%).
In the robotics domain, MLTI was particularly effective for complex tasks like "soccer" and "basketball" in Meta-World, where standard methods failed to show significant progress. The results suggest that weight-level priming is more powerful than simple input concatenation for complex adaptation. Furthermore, in the cognitive task, MLTI successfully mirrored human behavior: the model exhibited higher loss (a proxy for infant "looking time") when presented with inconsistent context-rule pairings, validating its potential as a model for cognitive control.
Conclusion
By integrating task-specific contextual information into the meta-learning process, MLTI provides a more biologically plausible and efficient approach to adaptation. While the current framework requires explicit context, future work could explore latent context discovery and scaling to larger architectures. Ultimately, this research demonstrates that guiding meta-learning with task information not only improves AI performance but also deepens our understanding of human cognitive flexibility.
Reference: Dubey, R., Grant, E., Luo, M., Narasimhan, K., & Griffiths, T. L. (2020). Connecting context-specific adaptation in humans to meta-learning. arXiv:2011.13782v2.
Summary updated