Omniglot Challenge Skill

The Lake-Salakhutdinov-Tenenbaum challenge for human-level concept learning through probabilistic program induction. Learn new concepts from a single example.

The Challenge

Five tasks that a single model must perform at human level:

ID level = Visual Turing Test identification rate (50% = indistinguishable from human)

Core Insight

"People learning new concepts can often generalize successfully from just a single example... We present a computational model that represents concepts as simple programs that best explain observed examples under a Bayesian criterion." — Lake, Salakhutdinov, Tenenbaum (Science, 2015)

Bayesian Program Learning (BPL)

Three key ingredients:

1. Compositionality: Concepts built from simpler primitives (strokes, parts)
2. Causality: Captures how data was actually generated (motor programs)
3. Learning to learn: Prior experience accelerates new concept acquisition

Character = Program(strokes, relations, noise)
P(concept | example) ∝ P(example | concept) × P(concept)

Dataset

• 50 alphabets from world's writing systems
• 1,623 characters total
• 20 examples per character (different writers)
• Stroke data included (not just images)

Splits

Why Current ML Falls Short

From the 3-year progress report (2019):

"Recent approaches are still far from human-like concept learning on Omniglot, a challenge that requires performing many tasks with a single model."

The Real Challenge

NOT just one-shot classification. The challenge is:

1. Single model for ALL five tasks
2. Minimal background training (5 alphabets, like humans)
3. Using stroke/motor program data, not just images
4. Compositionality and causality, not "learning from scratch"

References

• Lake, Salakhutdinov, Tenenbaum (2015). "Human-level concept learning through probabilistic program induction." Science 350:1332-1338. PDF

• Lake, Salakhutdinov, Tenenbaum (2019). "The Omniglot challenge: a 3-year progress report." Current Opinion in Behavioral Sciences 29:97-104. PDF

• Dataset: github.com/brendenlake/omniglot

Trit Assignment

• Trit: 0 (ERGODIC - coordinator)
• GF(3) Role: Bridges generative and discriminative approaches

Local Implementation

# ~/ies/worlding_skill_omniglot_entropy.py
from worlding_skill_omniglot_entropy import (
    ParallelOmniglotLearner,
    OmniglotCharacterFamily,
    BidirectionalCharacterLearner  # Read ↔ Write coupling
)

Connection to Active Inference

From Parr-Friston (Active Inference):

"Tenenbaum et al. (2006) established structure learning as a key objective in computational modeling and cognitive science."

BPL shares with Active Inference:

• Generative models of sensory data
• Hierarchical priors learned from experience
• Inference as explanation of observations

Key Quote

"Hofstadter famously argued that learning to recognize the characters in all the ways that people do contains most of the fundamental challenges of AI."

Gay.jl Colors (seed 2015)

#9858E7

#A81AA7

#BCD86F

#F283CD

#188DB2

Autopoietic Marginalia

The interaction IS the skill improving itself.

Every use of this skill is an opportunity for worlding:

• MEMORY (-1): Record what was learned
• REMEMBERING (0): Connect patterns to other skills
• WORLDING (+1): Evolve the skill based on use

Add Interaction Exemplars here as the skill is used.