Binz et al. highlight the potential of meta-learning to greatly enhance the flexibility of AI algorithms, as well as to approximate human behavior more accurately than traditional learning methods. We wish to emphasize a basic problem that lies underneath these two objectives, and in turn suggest another perspective of the required notion of “meta” in meta-learning: knowing what to learn.

Recent studies suggest that meta-learning may provide an original solution to an enduring puzzle about whether neural networks can explain compositionality – in particular, by raising the prospect that compositionality can be understood as an emergent property of an inner-loop learning algorithm. We elaborate on this hypothesis and consider its empirical predictions regarding the neural mechanisms and development of human compositionality.

Meta-learned models of cognition make optimal predictions for the actual stimuli presented to participants, but investigating judgment biases by constraining neural networks will be unwieldy. We suggest combining them with cognitive process models, which are more intuitive and explain biases. Rational process models, those that can sequentially sample from the posterior distributions produced by meta-learned models, seem a natural fit.

We propose that a principled understanding of meta-learning, as aimed for by the authors, benefits from linking the focus on learning with an equally strong focus on structure, which means to address the question: What are the meta-structures that can guide meta-learning?

Building on the affectivism approach, we expand on Binz et al.'s meta-learning research program by highlighting that emotion and other affective phenomena should be key to the modeling of human learning. We illustrate the added value of affective processes for models of learning across multiple domains with a focus on reinforcement learning, knowledge acquisition, and social learning.

Binz et al. argue that meta-learned models offer a new paradigm to study human cognition. Meta-learned models are proposed as alternatives to Bayesian models based on their capability to learn identical posterior predictive distributions. In our commentary, we highlight several arguments that reach beyond a predictive distribution-based comparison, offering new perspectives to evaluate the advantages of these modeling paradigms.

Meta-learning offers a promising framework to make sense of some parts of decision-making that have eluded satisfactory explanation. Here, we connect this research to work in animal behaviour and cognition in order to shed light on how and whether meta-learning could help us to understand the evolution of cognition.

We argue that the type of meta-learning proposed by Binz et al. generates models with low interpretability and falsifiability that have limited usefulness for neuroscience research. An alternative approach to meta-learning based on hyperparameter optimization obviates these concerns and can generate empirically testable hypotheses of biological computations.

This commentary examines the synergy between meta-learned models of cognition and integrative learning in enhancing animal and human learning outcomes. It highlights three integrative learning modes – holistic integration of parts, top-down reasoning, and generalization with in-depth analysis – and their alignment with meta-learned models of cognition. This convergence promises significant advances in educational practices, artificial intelligence, and cognitive neuroscience, offering a novel perspective on learning and cognition.

We are encouraged by the many positive commentaries on our target article. In this response, we recapitulate some of the points raised and identify synergies between them. We have arranged our response based on the tension between data and architecture that arises in the meta-learning framework. We additionally provide a short discussion that touches upon connections to foundation models.