Skip to content

The curriculum stack — automated training#

Curriculum handles automated shaping/training: moving an animal through increasingly demanding versions of a task based on its measured performance. Two roles:

  • aind-behavior-curriculum — an upstream dependency of aind-behavior-services. The framework's Task / TaskParameters subclass this package's base Task / Curriculum and reuse its SEMVER_REGEX. So the dependency order is aind-behavior-curriculum → aind-behavior-services.
  • Per-experiment curricula — e.g. aind-behavior-vr-foraging-curricula, which define the concrete training graphs for a paradigm.

The model#

A curriculum is a directed graph of stages:

  • Stage — each stage is a complete task config (an AindFooTaskLogic instance). Progressing the curriculum means swapping in a different, fully-specified task config — not running different code.
  • Transition — a predicate over performance metrics (rewards, choices, patch/site events) computed from the acquired dataset; when it evaluates true, the animal advances.
  • Policy — refines stages within or across sessions.

Curricula are semver'd and changed only via reviewed PRs, making training auditable: the exact stage graph that trained an animal is recorded and versioned.

Example: the VR Foraging depletion ladder#

learn-to-run → learn-to-stop → stochastic reward → multiple odors with depletion → graduation. Other VR Foraging curricula include depletion_stops_offset/rate, deterministic_reversals(_reward_capped), learning_sets, replenishment_depletion_offset, single_site, and template.

How it plugs into a run#

At launch, clabe's DefaultBehaviorPicker.pick_trainer_state(...) loads the animal's current curriculum TrainerState instead of a fixed task, and a CurriculumApp can compute the next-stage suggestion after the session using metrics derived from the acquired data. See running-an-experiment.

Citations#

  1. https://github.com/AllenNeuralDynamics/aind-behavior-curriculum
  2. vr-foraging-white-paper WHITEPAPER_DRAFT.md §7 (training)