Active inference goes to school
Notes on the DiPaolo-White-GueninCarlut-Constant-Clark paper
Paper link. Italic are my notes, or rephrasings of the paper under an AI perspective.
Abstract
How to incorporate disembodied AI into educational practices?
Active inference: human learning is epistemic foraging and prediction error minimization.
AI can be used in prepared learning environments for creating trajectories of self-correction.
Introduction
4E cogsci: embodied, embedded, enactive, extended. LLMs seem to undermine 4E. The jury is out: current disembodied AI systems do not seem to have true understanding, it is possible that hybrid LLM-robotics systems will, in the near future. Humans seem to ground meaning in dynamic sensorimotor interactions with the world.
In any case, how best to deal with LLMs has become a problem of literacy. The question in this paper: how can we leverage (disembodied) LLMs in (embodied and materially situated) human learning?
Active inference
Free energy principle: agents can control prediction error in two ways:
perceptual inference: updating their world models, and/or
actions to match the world with the model. For example, avoiding states with high prediction error, or actively creating a world to match the model.
World models are hierarchical, probabilistic, generative. Predicted future states are top-down, sensory information is bottom-up.
Precision (predicted uncertainty) both in predicted states and in actions.
Active inference is enacted and embodied
The Markov blanket, the veil of representation that separates the outside world from the inner screen, some say, is against the spirit of 4E.
Clark suggests that it is not the case: active inference is compatible with 4E, as modes of representations are profoundly action-oriented, actions do the cognitive work (procedural knowing!!), and thus active inference is the ideal framework to understand how the brain orchestrates embodied action.
The paper is agnostic about this, just uses active inference as a principled framework to explain action:
goal-oriented (change the world to match the model) - we call it exploitation in AI,
epistemic-oriented (collect data to update the model) - we call it exploration.
From an education point of view, what is important is error dynamics: error is valuable, it signals that learning to be done. There is a meta-optimization: it feels good to decrease prediction error faster than expected (and frustrating otherwise). Higher error can be good since it can be decreased, going through a well-designed trajectory of fast decrease of error is good. Csikszentmihalyi’s flow.
I’d add one more thing: probabilistic models predict states but also uncertainties. Going through a high-learning-rate trajectory on purpose requires that the uncertainty estimate is well-calibrated.
These states are not byproducts but crucial for action selection.
Important in non-stationary environments, when eg a resource can be exhausted, so error increases faster than expected, telling the agent to move on.
Explain curiosity and play: the pleasure of reducing uncertainty attracts us to just-uncertain-enough environments.
Learning in LLMs and active inference agents
Active inference learning is highly active: sensorimotor actions and (causal, generative) model development are driven by bodily needs towards affordances (food, water) to satisfy them. Meaningful engagement changes both the model and the agent/arena relationship.
LLMs are passive learners. Data selection is done by an outside agent. Text is relevant to whoever wrote it down, but not to the learning agent. Minimizing prediction errors is in fact similar to the active inference learner, and, in case of LLMs, the models are generative. The difference is that the data is not collected by the agent. The “feel” when interacting with LLMs is that they do not care about the truth, and it is actually not just a superficial observation. Since they have no stakes in their existence, they cannot care about importance and value.
Active inference goes to school
Can’t but notice the parallels between LLM training and traditional school, and active inference learning and Montessori. Might socially explain the widely-held belief that LLM training can lead to AGI. And ignored that before we go to school, we don’t get propositional training, we go through active inference learning.
Development (physical and biological growth → cognitive competences) and learning (refinement of cognitive competences and knowledge acquisition) are results of epistemic foraging, precision weighting (confidence in low-error policies), and error correction. Starts in the womb.
Active intervention (e.g. crying, grasping). Visual and other perceptual predictions. Surprise signals information gain. In classrooms: careful design of the materiality of learning.
Montessori has a long tradition of recognizing the material and spatial dimensions of learning.
Active inference learning in a Montessori (MM) classroom
Classical school: teacher-centered. (Active) teachers supply (propositional) knowledge, (passive) pupils absorb it, then teacher tests and grades. Similar to LLM training, although there is a hierarchical curriculum. Also: LLMs can more and more replace this type of teacher, except for the curriculum selection (importance of propositions).
Alternative school: (active) pupil engages with learning artifacts and ecologies, construct their own (perspectival → procedural → propositional), knowledge. MM: development and learning and embodied and enacted, unfolding in the classroom. Teacher’s role: organize the classroom as affordances.
Intrinsic motivation: no external rewards.
Attention to chosen meaningful actions.
Precision is obtained overcoming mistakes.
Classroom is a prepared environment with goals and affordances. Minds develop through active intervention. Exploration, curiosity, (not fantasy or make believe, Vervaeke’s serious) play: a form of epistemic foraging. Low stakes, the right amount of surprise, error tolerance. Embodied and affective drive to explore the boundaries. High uncertainty affords better-than-expected slope of error reduction. Playing with time scales: short term uncertainty → long term learning.
Experience first (perspectival), signify later (propositional). Solves the issue of assigning importance to propositions. Enhance errors on purpose, so the pupil learns how to detect and correct them.
MM pupils are good at error monitoring. Neutral stance to failure. Meta: learn to process errors without fear of uncertainty. The creative space. Mistakes are not a strong violation of expectations. Reminds me again to models that (meta-)model their uncertainties.
LLMs and the future of classroom
Hutto et al (2015): “You cannot directly teach anyone anything; at best, you can create activities that foster opportunities for a person to construct some targeted knowledge for themselves”.
Task of teacher: prepare and arrange a series of motives in a specially prepared environment. Educational material is not a set of unconnected external tools, but modes of material engagement. Hands and tools are made for action in action.
Active inference learning shows how these environments and interactions do their work:
Minimization of prediction errors by action,
the tuning of precision,
the role of affect and interoceptive feedback, and
the automatic shifting between exploit/explore dynamics.
LLMs can play a role in manufacturing the learning environment. In MM, teachers are mentors who help pupils learning by doing by themselves. LLMs can point out errors rather than autocorrect. Cognitive prosthetics, affording self-corrective and explorative behaviors. Reminds me of my conversation with Joel Gladd: LLM do “peer review”.