Source-linked AI summary
Meta-learning in natural and artificial intelligence
Jane X. Wang
TL;DR
The paper addresses how meta-learning research across AI, psychology, cognitive science, and neuroscience can be understood within a common framework. It reviews nested biological learning, connects it to AI models and Bayesian inference, and identifies structure learning and AI–neuroscience interaction as important future directions.
Problem
Research on learning to learn spans AI, psychology, cognitive science, and neuroscience, but these lines of work need a common meta-learning framework.
Method
The review recasts biological intelligence research through the lens of AI meta-learning and examines connections between neuroscience and AI.
Results
Meta-learning is ubiquitous across nested scales in nature, and neuroscience and cognitive science contain multiple related lines of work that connect with AI meta-learning.
Takeaways & Limitations
The complementary goals of AI and neuroscience create promising contact points for studying meta-learning, especially structure learning and biologically grounded learning systems.
Abstract
from arXiv · showhide
Meta-learning, or learning to learn, has gained renewed interest in recent years within the artificial intelligence community. However, meta-learning is incredibly prevalent within nature, has deep roots in cognitive science and psychology, and is currently studied in various forms within neuroscience. The aim of this review is to recast previous lines of research in the study of biological intelligence within the lens of meta-learning, placing these works into a common framework. More recent points of interaction between AI and neuroscience will be discussed, as well as interesting new directions that arise under this perspective.
1 Introduction
Meta-learning frames learning as becoming faster with experience through acquired inductive biases and knowledge. This review connects renewed AI interest with earlier psychological, cognitive, and neuroscientific research.
- 1 Introduction: Humans learn continuously across multiple timescales and levels of abstraction, with experience acquiring biases that make future learning more efficient.The review treats this process as meta-learning, or learning to learn.
- 1 Introduction: Deep learning systems achieve major successes but still require many orders of magnitude more data than humans.
- 1 Introduction: Recent AI work applies meta-learning to initial weights, update rules, input representations, and implicit learning algorithms.
- 1 Introduction: The review recasts psychological, cognitive, and neuroscientific research within contemporary AI meta-learning, emphasizing its prevalence and natural multiscale structure.It also identifies interactions between neuroscience and AI and proposes new research questions.
2 The scales of meta-learning: across and within lifetimes
Biological learning spans nested timescales, where longer-term learning can make shorter-term adaptation more efficient. Across development and lifetimes, innate structure and learned representations interact to support flexible learning.
- 2 The scales of meta-learning: across and within lifetimes: Biological learning adapts organisms to changing environmental challenges across mechanisms spanning different timescales.
- 2 The scales of meta-learning: across and within lifetimes: Longer-timescale learning can drive more efficient learning at shorter timescales, producing nested relationships across evolution, development, and behavior.
- 2 The scales of meta-learning: across and within lifetimes: The Baldwin effect describes how fast adaptation and learning can create positive selection pressure for genetic bases of those traits.Hinton and Nowlan demonstrated this selection for faster learning in simulation.
- 2 The scales of meta-learning: across and within lifetimes: Innate or developmentally predetermined behaviors interact with learned representations, as illustrated by innate place-cell propensity and environment-specific spatial content.
- 2 The scales of meta-learning: across and within lifetimes: Monkeys eventually learned a one-shot object-reward rule after extended experience with repeated object-role binding trials.
- 2 The scales of meta-learning: across and within lifetimes: Humans extend learning to learn through metacognition, metareasoning, educational learning, and hierarchical Bayesian models.
3 Neuroscience of meta-learning
Neuroscience research links meta-learning to regulation of learning algorithms, control over representations, structure learning, and Bayesian inference. These lines of work provide correspondences between biological learning and AI meta-learning.
- 3 Neuroscience of meta-learning: Neuroscience research connects meta-learning with learning control over existing representations, including schemas that support faster learning and memory integration.These processes are suggested to involve hippocampal-cortical interactions and memory consolidation.
- 3 Neuroscience of meta-learning: Biological meta-learning can regulate reinforcement-learning meta-parameters such as learning rate or discount factor through neuromodulators and anterior cingulate activity.
- 3 Neuroscience of meta-learning: Infants and adults learn latent structure, hierarchical rules, and statistical regularities, reflecting a bias toward structure learning.
- 3 Neuroscience of meta-learning: Bayesian probabilistic-program models explain few-example concept learning through structured, hierarchical priors.
- 3.3 Meta-learning as latent state and Bayesian inference: Neural networks trained across related tasks can perform hierarchical Bayesian inference by learning a prior that minimizes error on new related tasks.This correspondence also holds for model-agnostic meta-learning, where initial network parameters are meta-learned.
- 3.3 Meta-learning as latent state and Bayesian inference: In reinforcement learning, fast inner-loop learning and latent-state inference for decision-making or cognitive control are closely connected.
4 Bridging between AI and neuroscience: New questions and future directions
AI and neuroscience pursue different goals but can inform one another through meta-learning. Their interaction supports biologically grounded models and highlights structure learning and training dynamics as priorities for future research.
- 4 Bridging between AI and neuroscience: New questions and future directions: Neuroscience seeks to discover useful mental models already present in animals, whereas AI aims to engineer learning systems from scratch.
- 4 Bridging between AI and neuroscience: New questions and future directions: Weak inductive biases in randomly initialized deep networks shift the engineering problem toward constructing models of learning itself.
- 4 Bridging between AI and neuroscience: New questions and future directions: Deep reinforcement-learning models can capture animal-like metalearning effects consistent with previous neural findings.This interaction has encouraged work on more biologically plausible spiking networks and weight updates.
- 4 Bridging between AI and neuroscience: New questions and future directions: Future research should investigate how structure learning emerges, record neural signals during training, and characterize existing animal priors and their interactions with new learning.
5 Highlights
The review argues that multi-scale learning is ubiquitous in nature, connects cognitive and neuroscience research to meta-learning, and identifies complementary AI–neuroscience directions.
- Multiple scales of learning, and therefore meta-learning, are ubiquitous in nature.
- Existing neuroscience and cognitive science research addresses several aspects of meta-learning, including three lines emphasized by the review.
- AI and neuroscience pursue distinct but complementary goals, making meta-learning a promising point of contact.