← LearnWorkbench ▸·Path:

JUMPING FROG GENERATIVE MODEL LAB

A detailed educational sim game for the hierarchical generative model in the figure. Hidden states live inside the dotted circle, observations sit at the sensory boundary, and action changes tactile sensations. The simulator teaches the math of inference with a compact Bayesian teaching-world: posterior belief update from multiple sensory channels, precision weighting, modality-specific likelihoods, hidden-state priors, action contingencies, and cross-modal concept inference.
Standalone HTML
Generative Graph
Bayes Engine
Action Loop
Guided Lessons

Scenario Bay

Load a preset, then change hidden concepts, sensory observations, and action contingencies.

Inference Controls

lr beliefs
lr action
prior frog

Observation Panel

Each channel is binary in this teaching-world. The concept node predicts modality-specific percepts and sensations.

Observed sensations

Visual sensation
Auditory sensation
Interoceptive sensation
Tactile sensation

Action

Only tactile sensations are directly action-dependent here, following the dashed link in the figure.

Likelihood Builder

Edit how likely each observation is under the hidden concept “jumping frog”. Non-frog uses the complementary alternative.

Session Log

P(frog | data)
0.500
target > 0.90
Action
do_nothing
tactile-channel gain
g(action)
1.00
precision gain
LLR sum
0.000
Σ log-likelihood ratios
Best a (info)
—
expected surprise
Exact
0.500
single source of truth
GoalRaise P(frog | data) above 0.90 using multi-modal evidence.
WatchPosterior card above the canvas. Action scales tactile precision g(a), not the sampled value.
First moveClick Belief step. When posterior stalls, try Action step — it picks the action with the most expected info gain.

Status Deck

Posterior frog
0.500
Posterior non-frog
0.500
1 - frog
Visual evidence
0.000
log-like ratio
Audio evidence
0.000
log-like ratio
Intero evidence
0.000
log-like ratio
Tactile evidence
0.000
log-like ratio

Generative Graph

Bayes Engine

Action Loop

Guided Lessons

1. Hidden states and observations

Nodes inside the dotted circle are latent causes. Peripheral nodes are sensory data. Inference reverses the generative model.

2. One concept, many modalities

The concept “jumping frog” predicts visual, auditory, interoceptive, and tactile outcomes. Multisensory agreement strengthens the posterior.

3. Likelihoods

Each observation channel contributes a factor. The posterior is proportional to prior times the product of these factors.

4. Percepts

Beliefs about hidden states correspond to percepts tied to each modality: visual percept, auditory percept, tactile percept, and interoceptive percept.

5. Action contingencies

The dashed links show that action changes some sensory outcomes. Here action primarily changes tactile evidence.

6. Simplification

This teaching-world omits horizontal dependencies and deeper temporal structure, just like the figure’s simplification notes.

Layered Equation Panel

Beat Script

Pick a learning path

You can switch any time from the top bar. The math on the canvas is identical across paths.